Web v2: major visual upgrade, hamburger mobile menu, comments, SVG logo + Phase 11 novelty check

- Redesign: blur sticky header, hero gradient, elevated cards, footer grid
- Mobile: slide-down hamburger menu, full responsive pass
- Comments: persistent JSON storage (data/, excluded from deploys),
  honeypot + per-IP cooldown, public discussion page
- Logo: thesis-as-image (large model, small resident set) used as favicon
- research/novelty_check.md: components anticipated (QSpec, Lever, CATS,
  UGSD, SpecExec, mlx-flash); composition not found; 4 narrow surviving
  claims (N1-N4); attribution obligations recorded

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) parent 8e50d75

Showing 8 changed files with +658 and −159

added experiments/candidate_01/scale32b_w64.log +29 −0

		`@@ -0,0 +1,29 @@`
	1	+ Fetching 14 files: 0%\| \| 0/14 [00:00<?, ?it/s] Fetching 14 files: 100%\|██████████\| 14/14 [00:00<00:00, 5206.62it/s]
	2	+ Fetching 17 files: 0%\| \| 0/17 [00:00<?, ?it/s] Fetching 17 files: 100%\|██████████\| 17/17 [00:00<00:00, 6735.61it/s]
	3	+loading q4 resident …
	4	+q8 checkpoint (streamed): 34.8 GB
	5	+baseline: pure q4 …
	6	+ 1/12 (96 tok, 24.1 tok/s)
	7	+ 2/12 (96 tok, 26.6 tok/s)
	8	+ 3/12 (96 tok, 26.5 tok/s)
	9	+ 4/12 (96 tok, 26.7 tok/s)
	10	+ 5/12 (96 tok, 26.6 tok/s)
	11	+ 6/12 (96 tok, 26.7 tok/s)
	12	+ 7/12 (96 tok, 26.6 tok/s)
	13	+ 8/12 (96 tok, 26.5 tok/s)
	14	+ 9/12 (96 tok, 26.5 tok/s)
	15	+ 10/12 (96 tok, 26.6 tok/s)
	16	+ 11/12 (96 tok, 26.5 tok/s)
	17	+ 12/12 (96 tok, 26.3 tok/s)
	18	+runtime: mode=verify-all tau=2.0 W=64 …
	19	+ 1/12 (96 tok, 4 sweeps, 3 rollbacks, last sweep io 3.0s)
	20	+ 2/12 (96 tok, 10 sweeps, 9 rollbacks, last sweep io 2.9s)
	21	+ 3/12 (96 tok, 7 sweeps, 6 rollbacks, last sweep io 2.9s)
	22	+ 4/12 (96 tok, 4 sweeps, 3 rollbacks, last sweep io 3.0s)
	23	+ 5/12 (74 tok, 10 sweeps, 10 rollbacks, last sweep io 3.0s)
	24	+ 6/12 (96 tok, 15 sweeps, 15 rollbacks, last sweep io 2.8s)
	25	+ 7/12 (96 tok, 12 sweeps, 12 rollbacks, last sweep io 2.9s)
	26	+ 8/12 (85 tok, 11 sweeps, 11 rollbacks, last sweep io 3.0s)
	27	+ 9/12 (96 tok, 6 sweeps, 5 rollbacks, last sweep io 3.0s)
	28	+ 10/12 (95 tok, 11 sweeps, 11 rollbacks, last sweep io 3.0s)
	29	+ 11/12 (93 tok, 13 sweeps, 13 rollbacks, last sweep io 3.0s)

modified research/LOG.md +31 −0

		`@@ -309,3 +309,34 @@ Format per entry: date/time (local, with timezone) · question · experiment ·`
309	309	no window tuning.
310	310	- Decision: Phase 11 novelty check on the assembled architecture next; then throughput
311	311	levers or additional scale points.
	312	+
	313	+---
	314	+
	315	+## 2026-08-12 — Phase 11 novelty verification: composition survives narrowly; three components anticipated
	316	+
	317	+- Question: Charter §19 — is the candidate_01 architecture (resident q4 + margin-gated
	318	+ deferral + periodically streamed q8 verifier, bytes/token objective) already in the
	319	+ literature? Adversarial stance: assumed NOT novel.
	320	+- Experiment: Targeted web/arXiv/GitHub search per component and for the composition
	321	+ (~45 sources verified, key papers read from full PDFs). Deliverable:
	322	+ `research/novelty_check.md`.
	323	+- Result: Three 2025–2026 systems missed by the Phase 1 sweep materially shrink the
	324	+ claim: Lever (2605.16786 — DRAM-resident draft + flash-streamed target verifier,
	325	+ amortized per window, smartphones), CATS (2605.11186 — self-speculative shallow draft
	326	+ + verifier layers streamed from flash once per cycle, edge), U-HLM/UGSD
	327	+ (2412.12687 / 2603.11397 — draft-side uncertainty gates whether verification happens at
	328	+ all, cloud tier, incl. lossy top-R acceptance + rollback). QSpec/QuantSpec/ML-SpecQD own
	329	+ the shared-weights precision pair; mlx-flash owns MLX layer streaming
	330	+ (materialize→compute→release, every token).
	331	+- Interpretation: "Resident draft + streamed verifier" and "uncertainty-gated
	332	+ verification skipping" are ANTICIPATED as categories. Surviving as not-found: (N1)
	333	+ margin-gating the weight-streaming bytes of the same model's higher-precision variant
	334	+ on one machine; (N2) bytes-streamed-per-token vs checkpoint size as explicit objective,
	335	+ with τ as a runtime bytes↔quality knob; (N3) the training-free same-checkpoint q4/q8 pair
	336	+ as the draft/verify split of an out-of-core runtime; (N4, engineering) the unified-memory
	337	+ Metal instantiation with incremental verify-KV-cache. All public claims must use the
	338	+ attribution table and README language in novelty_check.md §3–§4.
	339	+- Decision: Proceed with narrowed claims only. Positioning for any writeup: extend
	340	+ Lever/CATS's regime with (a) precision-cast drafts (no training), (b) margin-priced
	341	+ verification I/O, (c) bytes/token accounting. Re-run this check before publication —
	342	+ prior-art density in this exact area is rising quarter over quarter.

added research/novelty_check.md +246 −0

		`@@ -0,0 +1,246 @@`
	1	+---
	2	+project: localvm-research
	3	+document: research/novelty_check
	4	+author: Simon-Pierre Boucher
	5	+contact: contact@spboucher.ai
	6	+created: 2026-08-12
	7	+status: draft
	8	+---
	9	+
	10	+# Phase 11 — Novelty verification of candidate_01 (margin-gated deferred refinement with a streamed verifier)
	11	+
	12	+Method. Per charter §19, the working assumption was **NOT NOVEL until evidence suggests
	13	+otherwise**. The search was adversarial: for each component of the architecture, and for the
	14	+composition, I searched for conceptual equivalents (not just identical terminology) across
	15	+arXiv (with emphasis on 2025–2026, i.e., after the Phase 1 sweep of 2026-08-11), ACL/EMNLP,
	16	+MLSys/OSDI/ASPLOS-adjacent venues, OpenReview, and the llama.cpp / MLX / mlx-lm GitHub
	17	+ecosystems. All sources are listed in §5 with access date 2026-08-12. The two closest prior
	18	+systems (Lever, CATS — both May 2026, both absent from the Phase 1 notes) were read from
	19	+their full PDFs, not abstracts.
	20	+
	21	+---
	22	+
	23	+## 1. The architecture under check (précis)
	24	+
	25	+A single-machine runtime for unified-memory consumer hardware (Apple Silicon, MLX/Metal):
	26	+
	27	+- A q4 quantization of the model is resident and decodes greedily at full speed,
	28	+ recording the top-1 logit margin of every token decision; **low-margin decisions are
	29	+ deferred** (margin mode) rather than escalated immediately.
	30	+- Every W tokens (W=32 measured), a **q8 variant of the same model — which does not fit in
	31	+ memory beside the base — is streamed layer-by-layer from NVMe**
	32	+ (materialize → compute → re-lazify; one strictly sequential pass over the checkpoint per
	33	+ sweep at the SSD device ceiling, ~11.6 GB/s) and teacher-forces the window in one pass
	34	+ against a persistent incremental q8 KV cache (O(window), not O(context)).
	35	+- On the first flip (checked at deferred positions only in margin mode; at all positions in
	36	+ verify-all mode) the runtime rolls back via KV-cache trim, accepts the q8 token, and
	37	+ resumes. Acceptance is deliberately approximate in margin mode (unchecked
	38	+ high-margin positions are trusted); verify-all is greedy-exact up to Metal numerics.
	39	+- Objective and headline metric: bytes streamed per generated token, decoupled from
	40	+ checkpoint size (9.4× under the q8 checkpoint at W=32 on Qwen3-32B); the margin threshold
	41	+ τ is a runtime knob trading bytes for quality (47%/83%/~100% of the q4→q8 quality gap at
	42	+ 150/211/237 MB/token on 1.7B; +0.28 nats over pure-q4 at 32B where q8 cannot run at all).
	43	+
	44	+---
	45	+
	46	+## 2. Per-component prior-art tables
	47	+
	48	+Verdict scale: ANTICIPATED (the component exists in print, differences are cosmetic) ·
	49	+PARTIALLY ANTICIPATED (the idea exists but in a materially different setting or without a
	50	+load-bearing property) · NOT FOUND (no conceptual equivalent located).
	51	+
	52	+### 2.1 Precision-tiered self-speculation (q4 draft / q8 verify, shared weights)
	53	+
	54	+\| Closest prior art \| What it does \| What it does NOT do \| Verdict \|
	55	+\|---\|---\|---\|---\|
	56	+\| QSpec (EMNLP 2025, arXiv 2410.11305) \| Same W4 weights; W4A4 draft, W4A16 verify; near-zero switching cost; lossless; 1.64× \| Both operand sets resident; verifies every token; objective is arithmetic speed, not capacity or bytes/token \| ANTICIPATED (core mechanism) \|
	57	+\| QuantSpec (Apple+Berkeley, ICML 2025) \| Self-speculation with 4-bit weights + 4-bit hierarchical KV draft, full-precision verify; >90% acceptance; edge-motivated \| Full-precision verifier fully resident and fully read; lossless only; latency objective \| ANTICIPATED (core mechanism) \|
	58	+\| ML-SpecQD (arXiv 2503.13565) \| MXFP4 direct cast of the target as turnkey draft (no training), 16-bit verify, >2× \| Same: verifier resident; verify-all; latency framing \| ANTICIPATED (core mechanism) \|
	59	+\| Sub-variant (a) verifier kept OUT of memory / streamed \| Lever and CATS stream their verifier from flash (see §2.3/2.4) — but their drafts are a separate small model (Lever) or a distilled shallow sub-network (CATS), never a precision cast of the same weights \| No published system pairs a low-bit resident cast with a streamed higher-precision variant of the same checkpoint \| NOT FOUND (within the precision-tier family) \|
	60	+\| Sub-variant (b) verify only uncertainty-flagged positions \| Exists outside this family (U-HLM, UGSD, SPRINTER — §2.2) \| No precision-tiered system gates which positions are verified \| NOT FOUND (within the precision-tier family) \|
	61	+\| Sub-variant (c) deliberately approximate acceptance \| Rich lossy-verification literature (§2.2): rank/top-R acceptance, judge heads, relaxed acceptance in TensorRT-LLM/SGLang; BiLD thresholds; taxonomy in arXiv 2607.26627 \| Not combined with precision tiers or a storage tier \| ANTICIPATED (in other settings) \|
	62	+\| Sub-variant (d) bytes-per-token objective \| SpecExec's economics imply bytes/token ÷ acceptance length; CATS names flash↔DRAM traffic as the bottleneck \| Nobody states or reports bytes-streamed-per-token vs checkpoint size as the optimization target and headline result \| NOT FOUND \|
	63	+
	64	+### 2.2 Confidence/margin-gated selective verification or escalation
	65	+
	66	+\| Closest prior art \| What it does \| What it does NOT do \| Verdict \|
	67	+\|---\|---\|---\|---\|
	68	+\| U-HLM (arXiv 2412.12687, IEEE 2025) — the sharpest hit \| On-device SLM drafts; tokens whose on-device uncertainty ≤ threshold skip uplink transmission and LLM verification entirely (74.8% of uplinks skipped); rejection-behavior predicted from draft-side uncertainty \| Uncertainty via temperature-perturbation consistency, not logit margin; expensive tier is a remote different model, cost is radio bytes, not weight-streaming I/O; no rollback of committed tokens; no storage tier \| PARTIALLY ANTICIPATED \|
	69	+\| UGSD (arXiv 2603.11397, Mar 2026) \| Edge draft; entropy-gated escalation of only high-uncertainty token blocks to a cloud verifier; top-R rank acceptance (lossy); suffix discarded on correction (rollback); adaptive block length; only 18.2% of tokens escalated \| Cloud tier (communication cost), different model as verifier, domain-specific (speech emotion captioning); no weight streaming; no bytes/token framing \| PARTIALLY ANTICIPATED \|
	70	+\| SPRINTER (arXiv 2502.04557) \| Tiny learned verifier predicts whether the target would accept; target LLM invoked only when a token is deemed unacceptable (approximate, not exact) \| Verifier is a learned classifier, not the draft's own margin; both models resident; no storage tier \| PARTIALLY ANTICIPATED \|
	71	+\| BiLD (NeurIPS 2023) \| Small model decodes; confidence-threshold fallback to large model + distance-threshold rollback of recent tokens \| Two separate resident models; no storage tier; no exactness mode; latency objective \| ANTICIPATED (gating + rollback pattern) \|
	72	+\| CALM (NeurIPS 2022) / cascades (FrugalGPT, RouteLLM) \| Margin/confidence-calibrated early exit with distribution-free risk control; request-level routing to expensive models \| Depth axis / request granularity; nothing about weight loading \| ANTICIPATED (signal + calibration recipe) \|
	73	+\| Token-level margin gating of a weight-loading (SSD-streaming) decision \| — \| Searched: "uncertainty triggered precision escalation", "confidence-gated weight loading", "margin-gated verification bytes", dynamic-precision line (PMPD, HOBBIT, DP-LLM — all feed-forward, no verification), lossy-verification taxonomy (2607.26627 — no memory-constrained verifier) \| NOT FOUND \|
	74	+
	75	+### 2.3 Layer-by-layer weight streaming from storage during a forward pass (unified memory)
	76	+
	77	+\| Closest prior art \| What it does \| What it does NOT do \| Verdict \|
	78	+\|---\|---\|---\|---\|
	79	+\| AirLLM (lyogavin/airllm) / ZeRO-Inference \| Layer-by-layer load→compute→free; capacity decoupling only \| Streams the whole model for every token; no draft; no selectivity \| ANTICIPATED (streaming primitive) \|
	80	+\| mlx-flash (GitHub, matt-k-wong) \| Exactly materialize→compute→release per layer in MLX on Apple Silicon (StreamingProxy wrapping, predictive I/O scheduler prefetching layer N+1 during layer N); "30B on 16 GB, 70B+ on 32 GB+" \| Streams on every token; no draft/verify role for the streamed model; no periodicity, no quality tiering \| ANTICIPATED (the MLX materialize→compute→re-lazify scheduling itself is not ours) \|
	81	+\| mlx-moe, vMLX, MLX issue #2878, MLX discussion #615 \| SSD expert streaming for MoE in MLX; open feature request for out-of-core streaming (Dec 2025); maintainer position that weights must be materialized/wired for GPU use \| Nothing combines a streamed model with a resident draft; #2878 does not mention speculation \| supports NOT FOUND for the verifier-role variant \|
	82	+\| Lever (arXiv 2605.16786, May 2026) \| DRAM-resident draft, flash-resident target streamed at each verification invocation on smartphones; "the SSM can remain resident in DRAM … the LLM can be stored in flash and invoked only intermittently"; I/O is 78–93% of verification latency (their Table 1) \| Separate small draft model (not same weights); CPU/NPU + UFS, not unified-memory Metal; verification invoked every speculative cycle (no margin-gated skipping); lossless greedy acceptance; latency objective \| ANTICIPATED (streamed-verifier-while-resident-draft as a category) \|
	83	+\| CATS (arXiv 2605.11186, May 2026) \| Self-speculative on edge: shallow sub-network drafts under the DRAM budget; intermediate layers "streamed from flash once per cycle" for shallow verification; remaining layers "offloaded from the flash chunk by chunk" for target verify; peak memory = target model alone; 5.08× \| Draft axis is depth (+ Reduced-KL distillation training), not a training-free precision cast; verification every cycle, not margin-deferred; no bytes/token metric; GPU/edge-Linux, not MLX/Metal \| ANTICIPATED (periodic flash-streamed verifier, window-amortized) \|
	84	+\| llama.cpp discussion #10466 (Nov 2024) \| Public proposal: fully-resident speculator + target layers ping-ponged into VRAM from CPU RAM for verification \| Never implemented; CPU-RAM/PCIe tier, not SSD; verify-all, exact \| corroborates ANTICIPATED (idea circulating) \|
	85	+
	86	+### 2.4 Windowed/batched deferred verification with rollback, where verifier cost is I/O
	87	+
	88	+\| Closest prior art \| What it does \| What it does NOT do \| Verdict \|
	89	+\|---\|---\|---\|---\|
	90	+\| SpecExec (NeurIPS 2024) \| Huge draft trees verified by a RAM-offloaded 70B in one pass; ~20 accepted tokens per full-model weight sweep; 10.6–18.7× over sequential offloaded decoding — the amortization economics of one weight sweep per window \| Separate 7B draft; CUDA/PCIe; exact sampling; never states bytes/token as metric \| ANTICIPATED (window-amortized I/O-bound verification) \|
	91	+\| Lever / CATS (above) \| Explicitly amortize flash reads of the verifier over the accepted window ("the cost of loading and executing the LLM is amortized across several output tokens" — Lever §1) \| See §2.3 \| ANTICIPATED \|
	92	+\| SpecOffload (arXiv 2505.10259) / SpecMoEOff (arXiv 2508.21706) / SpecInfer (ASPLOS 2024) / TriForce (COLM 2024) \| Speculation embedded into offloading pipelines (weights or KV) to hide transfer latency; TriForce: resident sparse-KV draft of the same weights, offloaded full KV verify \| Throughput/batch regimes; PCIe tiers; verify-all; exact \| ANTICIPATED \|
	93	+\| Rollback via KV-cache trim on first flip \| Standard speculative-decoding machinery; BiLD rollback policy; UGSD suffix discard \| — \| ANTICIPATED (not claimable at all) \|
	94	+\| Window-amortized verification whose cost is I/O and whose frequency/coverage is margin-gated \| — \| Lever/CATS/SpecExec verify every window; U-HLM/UGSD gate but have no weight I/O \| NOT FOUND \|
	95	+
	96	+### 2.5 The composition as a whole
	97	+
	98	+Search phrasings tried (plus synonyms): "resident draft streamed verifier", "quantized draft
	99	+full model verification SSD", "deferred verification language model rollback", "amortized
	100	+weight streaming verification", "two-tier quantization runtime offload verify", "speculative
	101	+decoding out-of-core verifier", "verifier offloading speculative decoding", "uncertainty
	102	+triggered precision escalation", "self-speculative decoding offloading edge", "flash-backed
	103	+target model speculative decoding", plus GitHub sweeps of llama.cpp / ik_llama.cpp / mlx /
	104	+mlx-lm issues and community MLX streaming projects.
	105	+
	106	+\| Candidate \| Overlap with our composition \| Decisive differences \|
	107	+\|---\|---\|---\|
	108	+\| CATS (May 2026) — nearest overall \| Resident cheap variant of the model drafts; verifier layers streamed from flash once per window; rollback; single memory-limited device; peak memory = one model \| Cheap variant is a depth truncation requiring distillation training, not a training-free precision cast; every cycle verified (no margin deferral; no τ knob; no approximate mode); metric is wall-clock speedup, not bytes/token; not unified-memory/Metal \|
	109	+\| Lever (May 2026) \| DRAM-resident draft + flash-streamed target verifier, invocation cost amortized over the accepted window; I/O identified as the dominant verification cost \| Separate draft model (no shared weights / no precision pair); verification never skipped by draft confidence; lossless; smartphone NPU/UFS \|
	110	+\| U-HLM + UGSD (Dec 2024 / Mar 2026) \| Draft-side uncertainty gates whether the expensive tier is invoked at all; UGSD adds block granularity, lossy top-R acceptance and suffix rollback \| Expensive tier is a remote, different model; the gated cost is communication, not weight-streaming bytes; nothing resides on or streams from storage \|
	111	+\| QSpec / QuantSpec / ML-SpecQD \| The q4-drafts / q8-verifies-with-shared-weights pair, proven acceptance >90% \| Verifier resident; verify-all; arithmetic-latency objective \|
	112	+
	113	+**Composition verdict: NOT FOUND as assembled — but only by the conjunction, and the margin
	114	+is thin.** Every load-bearing part exists separately, three of them in systems published
	115	+in 2025–2026 that the Phase 1 sweep missed (Lever, CATS, UGSD). Prior-art density around
	116	+this exact point is increasing quarter over quarter.
	117	+
	118	+---
	119	+
	120	+## 3. Honest bottom line — what survives, what must be attributed
	121	+
	122	+Must be attributed (do not claim):
	123	+
	124	+1. Precision-tiered draft/verify on shared weights — QSpec, QuantSpec, ML-SpecQD own it.
	125	+2. Resident draft + storage-streamed verifier, amortized over a token window — Lever
	126	+ and CATS own the category (with SpecExec owning the amortization economics and
	127	+ llama.cpp #10466 as a public 2024 proposal). Our verify-all mode is, at the category
	128	+ level, "Lever/CATS with a precision-cast draft on Apple Silicon."
	129	+3. Uncertainty-gated skipping of verification / escalation to an expensive tier — **U-HLM,
	130	+ UGSD, SPRINTER, with the older lineage BiLD / CALM / cascades**.
	131	+4. Deliberately lossy acceptance — the relaxed/lossy-verification literature (top-R rank
	132	+ acceptance in UGSD; judge/relaxed acceptance lines; taxonomy in arXiv 2607.26627).
	133	+5. Rollback via KV trim; teacher-forcing a window in one pass — standard speculative
	134	+ decoding machinery.
	135	+6. Layer-by-layer materialize→compute→release streaming in MLX on Apple Silicon —
	136	+ mlx-flash (and AirLLM/ZeRO-Inference generically). We did not invent the scheduling
	137	+ pattern; we gave the streamed pass a verifier role.
	138	+
	139	+Surviving as (narrowly) novel contributions:
	140	+
	141	+- *N1 — Margin-gated deferral of verification bytes.* Using the resident draft's top-1
	142	+ logit margin to decide **whether higher-precision weights are streamed from local storage
	143	+ at all** (which positions a sweep must check, and hence how much I/O the quality repair
	144	+ costs). Every located gating system gates compute or communication to a *different,
	145	+ remote or resident* model; none gates the weight-streaming I/O of the same model's
	146	+ higher-precision representation. NOT FOUND after targeted search.
	147	+- N2 — Bytes-streamed-per-token as the explicit objective, reported against checkpoint
	148	+ size (9.4× under at W=32), with the τ threshold exposed as a **runtime bytes↔quality
	149	+ knob** (47%→83%→~100% of the q4→q8 gap at 150→211→237 MB/token). Dynamic-precision
	150	+ systems (PMPD, HOBBIT, DP-LLM) have precision knobs without verification; lossy
	151	+ verification has quality knobs without a storage tier; no system prices quality in
	152	+ bytes-of-weights-streamed. NOT FOUND as a stated objective/metric.
	153	+- **N3 — The training-free same-checkpoint precision pair as the draft/verify split of an
	154	+ out-of-core runtime** (q4 resident / q8 streamed, no distillation, no auxiliary model),
	155	+ making "quality level that does not fit in RAM" usable on unified memory — demonstrated
	156	+ where the alternative (resident q8) is physically impossible (32B on 48 GB), judged
	157	+ +0.28 nats over the only configuration that fits. The conjunction is unoccupied; each
	158	+ half is not.
	159	+- N4 (engineering, weaker) — The unified-memory instantiation: strictly sequential
	160	+ sweep at SSD device ceiling, persistent incremental verify-KV-cache (O(window) sweeps)
	161	+ with trim-rollback, measured on Metal with the prefill/decode numerics ceiling
	162	+ identified. Claimable only as a measured system contribution on this platform, not as a
	163	+ concept.
	164	+
	165	+Claims that are now weaker than when candidate_01 was designed: "first
	166	+resident-draft/streamed-verifier runtime" (false — Lever, CATS), "first selective
	167	+verification by draft uncertainty" (false — U-HLM, SPRINTER, UGSD), "first
	168	+precision-pair self-speculation" (false — QSpec family), "first out-of-core amortized
	169	+verification" (false — SpecExec, Lever, CATS).
	170	+
	171	+---
	172	+
	173	+## 4. Recommended claim language for the README (defensible phrasing)
	174	+
	175	+> What is new here is the conjunction, and two specific properties of it. Precision-pair
	176	+> self-speculation (QSpec, QuantSpec, ML-SpecQD), storage-streamed verifiers behind a
	177	+> resident draft (Lever, CATS), window-amortized verification of an offloaded target
	178	+> (SpecExec), and uncertainty-gated escalation (BiLD, CALM, U-HLM, UGSD) all exist
	179	+> separately. To our knowledge, no prior system (i) uses the draft's own top-1 logit margin
	180	+> to gate how many bytes of higher-precision weights are streamed from local NVMe to
	181	+> repair quality — making the margin threshold τ a runtime knob that prices output quality
	182	+> in bytes-per-token — and (ii) does so with a **training-free precision pair of the same
	183	+> checkpoint** (4-bit resident, 8-bit streamed) on unified-memory hardware, so that a
	184	+> quality level whose weights cannot reside in RAM remains usable, with bytes streamed per
	185	+> token decoupled from checkpoint size (measured 9.4× under the streamed checkpoint at
	186	+> W=32). We explicitly do not claim novelty for precision-tiered drafting, streamed
	187	+> verification, windowed rollback, or confidence-gated escalation taken individually; the
	188	+> attribution table in `research/novelty_check.md` maps each component to its prior art.
	189	+
	190	+Avoid in all public text: "first", "novel runtime" unqualified, "no one has streamed a
	191	+verifier", "quality becomes a runtime variable" without the qualifier *via verification
	192	+bytes* (dynamic-precision papers already make quality a runtime variable without
	193	+verification).
	194	+
	195	+---
	196	+
	197	+## 5. Sources
	198	+
	199	+All accessed 2026-08-12.
	200	+
	201	+- QSpec: Speculative Decoding with Complementary Quantization Schemes (EMNLP 2025) — https://arxiv.org/abs/2410.11305 (accessed 2026-08-12)
	202	+- QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache (Apple, ICML 2025) — https://machinelearning.apple.com/research/quantspec (accessed 2026-08-12)
	203	+- ML-SpecQD: Multi-Level Speculative Decoding with Quantized Drafts — https://arxiv.org/pdf/2503.13565 (accessed 2026-08-12)
	204	+- Lever: Speculative LLM Inference on Smartphones (Wang et al., May 2026) — https://arxiv.org/pdf/2605.16786 (accessed 2026-08-12)
	205	+- CATS: Cascaded Adaptive Tree Speculation for Memory-Limited LLM Inference Acceleration (Han et al., May 2026) — https://arxiv.org/abs/2605.11186 (accessed 2026-08-12)
	206	+- CATS code repository — https://github.com/ElizaFuLan/CATS.git (accessed 2026-08-12, via paper)
	207	+- LLMCad: Fast and Scalable On-device Large Language Model Inference — https://arxiv.org/pdf/2309.04255 (accessed 2026-08-12)
	208	+- EdgeLLM: Fast On-Device LLM Inference With Speculative Decoding (IEEE TMC) — https://ieeexplore.ieee.org/document/10812936/ (accessed 2026-08-12)
	209	+- Dovetail: A CPU/GPU Heterogeneous Speculative Decoding for LLM inference (EMNLP 2025) — https://arxiv.org/abs/2412.18934 (accessed 2026-08-12)
	210	+- SpecOffload: Unlocking Latent GPU Capacity for LLM Inference on Resource-Constrained Devices — https://arxiv.org/html/2505.10259v2 (accessed 2026-08-12)
	211	+- Accelerating Mixture-of-Experts Inference by Hiding Offloading Latency with Speculative Decoding (SpecMoEOff) — https://arxiv.org/abs/2508.21706 (accessed 2026-08-12)
	212	+- SpecExec: Massively Parallel Speculative Decoding for Interactive LLM Inference on Consumer Devices (NeurIPS 2024) — https://arxiv.org/html/2406.02532v1 (accessed 2026-08-12, re-verified)
	213	+- SpecInfer: Accelerating LLM Serving with Tree-based Speculative Inference and Verification (ASPLOS 2024) — https://arxiv.org/abs/2305.09781 (accessed 2026-08-12, re-verified)
	214	+- TriForce: Lossless Acceleration of Long Sequence Generation with Hierarchical Speculative Decoding (COLM 2024) — https://arxiv.org/abs/2404.11912 (accessed 2026-08-12)
	215	+- MemSpec: Memory-Aware Runtime for Adaptive Draft Scheduling in Speculative Decoding on Edge Devices (LCTES 2026) — https://doi.org/10.1145/3814943.3816174 (accessed 2026-08-12)
	216	+- HiSpec: Hierarchical Speculative Decoding for LLMs — https://arxiv.org/abs/2510.01336 (accessed 2026-08-12)
	217	+- SPRINTER: Speeding up Speculative Decoding via Sequential Approximate Verification — https://arxiv.org/abs/2502.04557 (accessed 2026-08-12)
	218	+- U-HLM: Uncertainty-Aware Hybrid Inference with On-Device Small and Remote Large Language Models — https://arxiv.org/abs/2412.12687 (accessed 2026-08-12)
	219	+- Communication-Efficient Hybrid Language Model via Uncertainty-Aware Opportunistic and Compressed Transmission — https://arxiv.org/pdf/2505.11788 (accessed 2026-08-12)
	220	+- Edge–Cloud Collaborative Speech Emotion Captioning via Token-Level Speculative Decoding (UGSD, Mar 2026) — https://arxiv.org/html/2603.11397v1 (accessed 2026-08-12)
	221	+- Quantize-Sample-and-Verify: LLM Acceleration via Adaptive Edge-Cloud Speculative Decoding — https://arxiv.org/pdf/2507.00605 (accessed 2026-08-12)
	222	+- Fast and Cost-effective Speculative Edge-Cloud Decoding with Early Exits — https://arxiv.org/abs/2505.21594 (accessed 2026-08-12)
	223	+- DSD: A Distributed Speculative Decoding Solution for Edge-Cloud Agile Large Model Serving — https://arxiv.org/pdf/2511.21669 (accessed 2026-08-12)
	224	+- Speculative Decoding with Big Little Decoder (BiLD, NeurIPS 2023) — https://arxiv.org/abs/2302.07863 (accessed 2026-08-12, re-verified)
	225	+- Confident Adaptive Language Modeling (CALM, NeurIPS 2022) — https://arxiv.org/abs/2207.07061 (accessed 2026-08-12, re-verified)
	226	+- Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes (Jul 2026) — https://arxiv.org/abs/2607.26627 (accessed 2026-08-12)
	227	+- Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match (FLy) — https://arxiv.org/pdf/2511.22972 (accessed 2026-08-12)
	228	+- A Practical Investigation of Training-free Relaxed Speculative Decoding — https://arxiv.org/html/2607.08690 (accessed 2026-08-12)
	229	+- Spec-VLA: Speculative Decoding for Vision-Language-Action Models with Relaxed Acceptance — https://arxiv.org/html/2507.22424v1 (accessed 2026-08-12)
	230	+- When Is a Draft Accepted? A Theory of Acceptance in Speculative Decoding — https://arxiv.org/html/2606.30265 (accessed 2026-08-12)
	231	+- Statistically-Lossless Quantization of Large Language Models (EAR metric; static, no runtime escalation) — https://arxiv.org/pdf/2605.02404 (accessed 2026-08-12)
	232	+- Mixed-Precision Quantization for Language Models: Techniques and Prospects (incl. HOBBIT precision-on-miss loading) — https://arxiv.org/html/2510.16805v1 (accessed 2026-08-12)
	233	+- Progressive Mixed-Precision Decoding for Efficient LLM Inference (PMPD) — https://arxiv.org/pdf/2410.13461 (accessed 2026-08-12, re-verified)
	234	+- Lossless but Not Free: An Empirical Anatomy of Speculative Decoding on Consumer Hardware (Jul 2026) — https://arxiv.org/html/2607.17283v1 (accessed 2026-08-12)
	235	+- Speculative Speculative Decoding (ICLR 2026) — https://arxiv.org/pdf/2603.03251 (accessed 2026-08-12)
	236	+- llama.cpp discussion #10466: Speculative decoding potential for running big LLMs on consumer grade GPUs (resident speculator + layer ping-pong proposal, unimplemented) — https://github.com/ggml-org/llama.cpp/discussions/10466 (accessed 2026-08-12)
	237	+- llama.cpp issue #21453: Research: Speculative Decoding for Low-Latency CPU Inference — https://github.com/ggml-org/llama.cpp/issues/21453 (accessed 2026-08-12)
	238	+- ik_llama.cpp discussion #839: speculative decoding via network — https://github.com/ikawrakow/ik_llama.cpp/discussions/839 (accessed 2026-08-12)
	239	+- mlx-flash: Flash weight streaming for MLX (per-layer StreamingProxy, predictive I/O, every-token streaming, no draft/verify) — https://github.com/matt-k-wong/mlx-flash (accessed 2026-08-12)
	240	+- MLX issue #2878: On-disk (out-of-core) streaming of model weights for low-RAM inference (Dec 2025; no speculation proposed) — https://github.com/ml-explore/mlx/issues/2878 (accessed 2026-08-12)
	241	+- mlx-moe: router-selected expert streaming from SSD with MLX — https://github.com/mu-hashmi/mlx-moe (accessed 2026-08-12)
	242	+- vMLX: SSD expert streaming + disk cache for MoE on Apple Silicon — https://github.com/jjang-ai/vmlx (accessed 2026-08-12)
	243	+- dflash-mlx: Lossless DFlash speculative decoding for MLX (adaptive verify block, KV SSD spill; target resident) — https://github.com/bstnxbt/dflash-mlx (accessed 2026-08-12)
	244	+- MLX discussion #615: Loading models with mmap (0.025 tok/s overflow result) — https://github.com/ml-explore/mlx/discussions/615 (accessed 2026-08-12, re-verified)
	245	+- LLM in a flash: Efficient LLM Inference with Limited Memory (Apple) — https://arxiv.org/abs/2312.11514 (accessed 2026-08-12, re-verified)
	246	+- AirLLM (layer-by-layer streaming baseline) — https://github.com/lyogavin/airllm (accessed 2026-08-12)

modified web/lib/render.js +27 −6

		`@@ -132,6 +132,7 @@ const NAV = [`
132	132	["/results", "Results"],
133	133	["/code", "Code"],
134	134	["/doc/research/bibliography.md", "Bibliography"],
	135	+ ["/comments", "Comments"],
135	136	["/about", "About"],
136	137	];
137	138
		`@@ -150,21 +151,41 @@ function layout({ title, active, body, buildInfo = {} }) {`
150	151	<title>${esc(title)} · localvm-research</title>
151	152	<meta name="description" content="localvm-research — running LLMs larger than memory on consumer Apple Silicon. Research by Simon-Pierre Boucher.">
152	153	<link rel="stylesheet" href="/static/style.css">
153		−<link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><text y='0.9em' font-size='90'>🧠</text></svg>">
	154	+<link rel="icon" type="image/svg+xml" href="/static/logo.svg">
	155	+<link rel="apple-touch-icon" href="/static/logo.svg">
154	156	</head>
155	157	<body>
156	158	<header class="site-header">
157	159	<div class="wrap header-row">
158		− <a class="brand" href="/">localvm<span class="brand-dim">-research</span></a>
159		− <nav class="site-nav">${nav}</nav>
	160	+ <a class="brand" href="/"><img class="brand-logo" src="/static/logo.svg" alt="" width="26" height="26">localvm<span class="brand-dim">-research</span></a>
	161	+ <button class="nav-toggle" id="nav-toggle" aria-label="Menu" aria-expanded="false">
	162	+ <span></span><span></span><span></span>
	163	+ </button>
	164	+ <nav class="site-nav" id="site-nav">${nav}</nav>
160	165	</div>
161	166	</header>
162	167	<main class="wrap">${body}</main>
163	168	<footer class="site-footer">
164		− <div class="wrap">
165		− <span>© 2026 Simon-Pierre Boucher — <a href="mailto:contact@spboucher.ai">contact@spboucher.ai</a> · All rights reserved (research code)</span>
166		− <span class="footer-meta">snapshot ${esc(commit)}${synced ? " · synced " + esc(synced) : ""} · Apple M5 Max / 48 GB / macOS 27</span>
	169	+ <div class="wrap footer-grid">
	170	+ <div>
	171	+ <span class="footer-brand">localvm-research</span>
	172	+ <p class="footer-tag">Running LLMs larger than memory on consumer Apple Silicon —
	173	+ an open, falsification-driven research trail.</p>
	174	+ </div>
	175	+ <div class="footer-col">
	176	+ <span class="footer-h">Project</span>
	177	+ <a href="/doc/CLAUDE.md">Charter</a>
	178	+ <a href="/doc/research/LOG.md">Research log</a>
	179	+ <a href="/experiments">Experiments</a>
	180	+ </div>
	181	+ <div class="footer-col">
	182	+ <span class="footer-h">Author</span>
	183	+ <span>Simon-Pierre Boucher</span>
	184	+ <a href="mailto:contact@spboucher.ai">contact@spboucher.ai</a>
	185	+ <span class="footer-meta">snapshot ${esc(commit)}${synced ? " · " + esc(synced) : ""}</span>
	186	+ </div>
167	187	</div>
	188	+ <div class="wrap footer-legal">© 2026 Simon-Pierre Boucher · All rights reserved (research code) · Apple M5 Max / 48 GB / macOS 27</div>
168	189	</footer>
169	190	<script src="/static/app.js" defer></script>
170	191	</body>

modified web/public/app.js +19 −0

		`@@ -11,6 +11,25 @@`
11	11	// ============================================================================
12	12	"use strict";
13	13
	14	+// mobile hamburger menu
	15	+(function () {
	16	+ const toggle = document.getElementById("nav-toggle");
	17	+ const nav = document.getElementById("site-nav");
	18	+ if (!toggle \|\| !nav) return;
	19	+ toggle.addEventListener("click", () => {
	20	+ const open = nav.classList.toggle("open");
	21	+ toggle.classList.toggle("open", open);
	22	+ toggle.setAttribute("aria-expanded", String(open));
	23	+ });
	24	+ nav.addEventListener("click", (e) => {
	25	+ if (e.target.tagName === "A") {
	26	+ nav.classList.remove("open");
	27	+ toggle.classList.remove("open");
	28	+ }
	29	+ });
	30	+})();
	31	+
	32	+// chart hover tooltips
14	33	(function () {
15	34	const fig = document.querySelector(".chart-fig");
16	35	if (!fig) return;

added web/public/logo.svg +17 −0

		`@@ -0,0 +1,17 @@`
	1	+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" role="img" aria-label="localvm-research logo">
	2	+ <!-- ==========================================================================
	3	+ Project : localvm-research \| File: web/public/logo.svg
	4	+ Purpose : Logo/favicon — a large model (outline) with a small resident
	5	+ working set (solid) and streamed blocks (light), the project thesis
	6	+ Author : Simon-Pierre Boucher \| Contact : contact@spboucher.ai
	7	+ Created : 2026-08-12 \| Modified : 2026-08-12 \| License: All rights reserved
	8	+ =========================================================================== -->
	9	+ <rect x="4" y="4" width="56" height="56" rx="14" fill="#fcfcfb"/>
	10	+ <rect x="6" y="6" width="52" height="52" rx="12" fill="#edf3fb" stroke="#2a78d6" stroke-width="3.5"/>
	11	+ <!-- streamed (cold) blocks -->
	12	+ <rect x="34" y="14" width="16" height="16" rx="4.5" fill="#9ec5f4"/>
	13	+ <rect x="34" y="34" width="16" height="16" rx="4.5" fill="#cde2fb"/>
	14	+ <rect x="14" y="14" width="16" height="16" rx="4.5" fill="#cde2fb"/>
	15	+ <!-- resident working set -->
	16	+ <rect x="14" y="34" width="16" height="16" rx="4.5" fill="#2a78d6"/>
	17	+</svg>

modified web/public/style.css +221 −153

		`@@ -1,7 +1,7 @@`
1	1	/* ============================================================================
2	2	* Project : localvm-research
3	3	* File : web/public/style.css
4		− * Purpose : Light-theme styling for the research showcase platform
	4	+ * Purpose : Light-theme styling for the research showcase platform (v2)
5	5	* Author : Simon-Pierre Boucher
6	6	* Contact : contact@spboucher.ai
7	7	* Created : 2026-08-12
		`@@ -22,157 +22,239 @@`
22	22	--border: rgba(11, 11, 11, 0.10);
23	23	--accent: #2a78d6;
24	24	--accent-dark: #1c5cab;
	25	+ --accent-wash: #edf3fb;
25	26	--good: #0ca30c;
26	27	--good-text: #006300;
27		− --warn: #eda100;
	28	+ --radius: 12px;
	29	+ --shadow-1: 0 1px 2px rgba(11,11,11,0.04), 0 1px 3px rgba(11,11,11,0.05);
	30	+ --shadow-2: 0 4px 14px rgba(11,11,11,0.07), 0 2px 4px rgba(11,11,11,0.05);
28	31	}
29	32
30	33	* { box-sizing: border-box; }
31		−html { -webkit-text-size-adjust: 100%; }
	34	+html { -webkit-text-size-adjust: 100%; scroll-behavior: smooth; scroll-padding-top: 76px; }
32	35	body {
33	36	margin: 0;
34	37	font-family: system-ui, -apple-system, "Segoe UI", sans-serif;
35	38	background: var(--page);
36	39	color: var(--ink);
37		− line-height: 1.6;
	40	+ line-height: 1.65;
38	41	font-size: 16px;
39	42	}
40		−.wrap { max-width: 1080px; margin: 0 auto; padding: 0 24px; }
	43	+.wrap { max-width: 1120px; margin: 0 auto; padding: 0 24px; }
41	44	a { color: var(--accent-dark); text-decoration: none; }
42	45	a:hover { text-decoration: underline; }
43	46	code, pre, .mono-small { font-family: ui-monospace, SFMono-Regular, Menlo, monospace; }
	47	+::selection { background: #cde2fb; }
44	48
45		−/* header */
	49	+/* ---------------------------------------------------------------- header */
46	50	.site-header {
47		− background: var(--surface);
	51	+ position: sticky; top: 0; z-index: 40;
	52	+ background: rgba(252, 252, 251, 0.85);
	53	+ -webkit-backdrop-filter: saturate(180%) blur(14px);
	54	+ backdrop-filter: saturate(180%) blur(14px);
48	55	border-bottom: 1px solid var(--grid);
49		− position: sticky; top: 0; z-index: 10;
50	56	}
51		−.header-row { display: flex; align-items: center; justify-content: space-between; height: 56px; }
52		−.brand { font-weight: 700; font-size: 18px; color: var(--ink); letter-spacing: -0.02em; }
	57	+.header-row { display: flex; align-items: center; justify-content: space-between; height: 60px; gap: 16px; }
	58	+.brand {
	59	+ font-weight: 750; font-size: 18.5px; color: var(--ink); letter-spacing: -0.02em;
	60	+ display: flex; align-items: center; gap: 8px; white-space: nowrap;
	61	+}
	62	+.brand-logo { display: block; border-radius: 7px; }
53	63	.brand-dim { color: var(--muted); font-weight: 500; }
54		−.site-nav { display: flex; gap: 4px; flex-wrap: wrap; }
	64	+.site-nav { display: flex; gap: 2px; flex-wrap: wrap; }
55	65	.site-nav a {
56		− padding: 6px 12px; border-radius: 6px; color: var(--ink-2); font-size: 14px; font-weight: 500;
	66	+ padding: 7px 13px; border-radius: 8px; color: var(--ink-2);
	67	+ font-size: 14px; font-weight: 550;
	68	+ transition: background 0.15s, color 0.15s;
57	69	}
58		−.site-nav a:hover { background: var(--page); text-decoration: none; color: var(--ink); }
59		−.site-nav a.active { color: var(--accent-dark); background: #edf3fb; }
	70	+.site-nav a:hover { background: var(--accent-wash); text-decoration: none; color: var(--accent-dark); }
	71	+.site-nav a.active { color: var(--accent-dark); background: var(--accent-wash); font-weight: 650; }
	72	+.nav-toggle { display: none; }
60	73
61		−/* hero */
62		−.hero { padding: 56px 0 8px; }
	74	+/* ---------------------------------------------------------------- hero */
	75	+.hero {
	76	+ padding: 72px 0 16px;
	77	+ background:
	78	+ radial-gradient(1000px 380px at 15% -20%, rgba(42,120,214,0.10), transparent 65%),
	79	+ radial-gradient(800px 300px at 90% -30%, rgba(27,175,122,0.07), transparent 60%);
	80	+ margin: 0 -24px; padding-left: 24px; padding-right: 24px;
	81	+}
63	82	.kicker {
64		− text-transform: uppercase; letter-spacing: 0.08em; font-size: 12px; font-weight: 600;
65		− color: var(--accent-dark); margin: 0 0 12px;
	83	+ text-transform: uppercase; letter-spacing: 0.1em; font-size: 12px; font-weight: 700;
	84	+ color: var(--accent-dark); margin: 0 0 14px;
	85	+ display: inline-flex; align-items: center; gap: 8px;
66	86	}
67		−.hero h1 { font-size: 40px; line-height: 1.15; letter-spacing: -0.02em; margin: 0 0 16px; }
68		−.lede { font-size: 18px; color: var(--ink-2); max-width: 760px; margin: 0 0 12px; }
69		−.lede-small { color: var(--ink-2); max-width: 760px; }
	87	+.kicker::before { content: ""; width: 22px; height: 2px; background: var(--accent); border-radius: 2px; }
	88	+.hero h1 { font-size: 46px; line-height: 1.08; letter-spacing: -0.028em; margin: 0 0 18px; font-weight: 780; }
	89	+.lede { font-size: 18.5px; color: var(--ink-2); max-width: 780px; margin: 0 0 14px; }
	90	+.lede-small { color: var(--ink-2); max-width: 780px; }
70	91	.lede-eq code {
71	92	font-size: 13px; background: var(--surface); border: 1px solid var(--grid);
72		− border-radius: 6px; padding: 6px 10px; color: var(--ink-2); display: inline-block;
	93	+ border-radius: 8px; padding: 8px 14px; color: var(--ink-2); display: inline-block;
	94	+ box-shadow: var(--shadow-1);
73	95	}
74	96
75		−/* stat tiles */
76		−.tiles { display: grid; grid-template-columns: repeat(auto-fit, minmax(180px, 1fr)); gap: 16px; margin: 32px 0; }
	97	+/* ---------------------------------------------------------------- tiles */
	98	+.tiles { display: grid; grid-template-columns: repeat(auto-fit, minmax(190px, 1fr)); gap: 16px; margin: 36px 0; }
77	99	.tile {
78		− background: var(--surface); border: 1px solid var(--grid); border-radius: 10px;
79		− padding: 20px 22px; display: flex; flex-direction: column; gap: 2px;
	100	+ background: var(--surface); border: 1px solid var(--grid); border-radius: var(--radius);
	101	+ padding: 22px 24px; display: flex; flex-direction: column; gap: 2px;
	102	+ box-shadow: var(--shadow-1); position: relative; overflow: hidden;
	103	+}
	104	+.tile::before {
	105	+ content: ""; position: absolute; left: 0; top: 0; bottom: 0; width: 3px;
	106	+ background: linear-gradient(180deg, var(--accent), #6da7ec);
80	107	}
81		−.tile-value { font-size: 34px; font-weight: 700; letter-spacing: -0.02em; }
	108	+.tile-value { font-size: 36px; font-weight: 760; letter-spacing: -0.025em; line-height: 1.1; }
82	109	.tile-denom { font-size: 20px; color: var(--muted); font-weight: 500; }
83		−.tile-label { color: var(--ink-2); font-size: 13.5px; }
	110	+.tile-label { color: var(--ink-2); font-size: 13.5px; margin-top: 4px; }
84	111
85		−/* layout blocks */
	112	+/* ---------------------------------------------------------------- cards */
86	113	.split { display: grid; grid-template-columns: 3fr 2fr; gap: 20px; margin: 8px 0 20px; }
87		−@media (max-width: 880px) { .split { grid-template-columns: 1fr; } .hero h1 { font-size: 30px; } }
88	114	.card {
89		− background: var(--surface); border: 1px solid var(--grid); border-radius: 10px;
90		− padding: 24px 26px; margin-bottom: 20px;
	115	+ background: var(--surface); border: 1px solid var(--grid); border-radius: var(--radius);
	116	+ padding: 26px 28px; margin-bottom: 20px; box-shadow: var(--shadow-1);
	117	+}
	118	+.card h2 { margin-top: 0; font-size: 20px; letter-spacing: -0.015em; font-weight: 700; }
	119	+.more { font-weight: 650; font-size: 14px; }
	120	+.card-head-row { display: flex; justify-content: space-between; align-items: baseline; gap: 12px; flex-wrap: wrap; }
	121	+.btn {
	122	+ display: inline-block; background: var(--accent); color: #fff; border: 0;
	123	+ padding: 10px 20px; border-radius: 9px; font-size: 14.5px; font-weight: 650;
	124	+ cursor: pointer; transition: background 0.15s, transform 0.1s; font-family: inherit;
91	125	}
92		−.card h2 { margin-top: 0; font-size: 20px; letter-spacing: -0.01em; }
93		−.more { font-weight: 600; font-size: 14px; }
	126	+.btn:hover { background: var(--accent-dark); }
	127	+.btn:active { transform: translateY(1px); }
94	128
95		−/* phases */
	129	+/* ---------------------------------------------------------------- phases */
96	130	.phases { list-style: none; margin: 0; padding: 0; }
97		−.phase { display: flex; align-items: flex-start; gap: 12px; padding: 10px 0; border-bottom: 1px solid var(--grid); }
	131	+.phase { display: flex; align-items: flex-start; gap: 12px; padding: 11px 0; border-bottom: 1px solid var(--grid); }
98	132	.phase:last-child { border-bottom: 0; }
99	133	.phase strong { display: block; font-size: 14.5px; }
100	134	.phase-detail { font-size: 13px; color: var(--muted); }
101	135	.phase-dot { width: 10px; height: 10px; border-radius: 50%; margin-top: 7px; flex: 0 0 auto; background: var(--baseline); }
102		−.phase.done .phase-dot { background: var(--good); }
103		−.phase.in-progress .phase-dot { background: var(--accent); }
	136	+.phase.done .phase-dot { background: var(--good); box-shadow: 0 0 0 3px rgba(12,163,12,0.14); }
	137	+.phase.in-progress .phase-dot { background: var(--accent); box-shadow: 0 0 0 3px rgba(42,120,214,0.16); }
104	138	.phase .badge { margin-left: auto; }
105	139	.badge {
106		− font-size: 11.5px; font-weight: 600; padding: 3px 9px; border-radius: 20px;
	140	+ font-size: 11.5px; font-weight: 650; padding: 3px 10px; border-radius: 20px;
107	141	white-space: nowrap; align-self: center; border: 1px solid var(--grid); color: var(--ink-2);
108	142	}
109	143	.badge-done { color: var(--good-text); background: #eef7ee; border-color: #cfe8cf; }
110		−.badge-in-progress, .badge-has-results { color: var(--accent-dark); background: #edf3fb; border-color: #cfe0f5; }
	144	+.badge-in-progress, .badge-has-results { color: var(--accent-dark); background: var(--accent-wash); border-color: #cfe0f5; }
111	145	.badge-pending, .badge-scaffolded { color: var(--muted); background: var(--page); }
112	146
113		−/* doc cards */
114		−.page-title { font-size: 30px; letter-spacing: -0.02em; margin: 40px 0 8px; }
115		−.section-title { font-size: 20px; margin: 32px 0 8px; }
116		−.doc-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(280px, 1fr)); gap: 16px; margin: 16px 0; }
	147	+/* ---------------------------------------------------------------- doc cards */
	148	+.page-title { font-size: 32px; letter-spacing: -0.025em; margin: 44px 0 10px; font-weight: 760; }
	149	+.section-title { font-size: 20px; margin: 34px 0 10px; font-weight: 700; }
	150	+.doc-grid { display: grid; grid-template-columns: repeat(auto-fill, minmax(290px, 1fr)); gap: 16px; margin: 18px 0; }
117	151	.doc-card {
118		− background: var(--surface); border: 1px solid var(--grid); border-radius: 10px;
119		− padding: 18px 20px; color: var(--ink); display: block;
	152	+ background: var(--surface); border: 1px solid var(--grid); border-radius: var(--radius);
	153	+ padding: 20px 22px; color: var(--ink); display: block; box-shadow: var(--shadow-1);
	154	+ transition: box-shadow 0.18s, transform 0.18s, border-color 0.18s;
120	155	}
121		−.doc-card:hover { border-color: var(--accent); text-decoration: none; }
	156	+.doc-card:hover { border-color: #b9d2f0; text-decoration: none; box-shadow: var(--shadow-2); transform: translateY(-2px); }
122	157	.doc-card.disabled { opacity: 0.55; pointer-events: none; }
123		−.doc-card h3 { margin: 0 0 6px; font-size: 16px; }
124		−.doc-card p { margin: 0 0 8px; font-size: 13.5px; color: var(--ink-2); }
	158	+.doc-card h3 { margin: 0 0 6px; font-size: 16px; font-weight: 680; }
	159	+.doc-card p { margin: 0 0 10px; font-size: 13.5px; color: var(--ink-2); }
125	160	.runs { font-size: 12px; color: var(--muted); margin-left: 8px; }
126	161	.note-line { color: var(--ink-2); font-size: 14px; }
127	162
128		−/* markdown body */
	163	+/* ---------------------------------------------------------------- doc view */
129	164	.doc { max-width: 860px; margin: 32px auto; }
130		−.crumb { font-size: 13px; color: var(--muted); margin: 24px 0 4px; }
	165	+.crumb { font-size: 13px; color: var(--muted); margin: 26px 0 4px; }
131	166	.chips { display: flex; flex-wrap: wrap; gap: 8px; margin: 10px 0 18px; }
132	167	.chip {
133	168	font-size: 12px; background: var(--surface); border: 1px solid var(--grid);
134	169	border-radius: 20px; padding: 3px 12px; color: var(--ink-2);
135	170	}
136	171	.chip-k { color: var(--muted); margin-right: 6px; text-transform: uppercase; font-size: 10px; letter-spacing: 0.05em; }
	172	+.doc-head { margin: 30px 0 6px; }
	173	+.doc-head-row { display: flex; align-items: center; gap: 12px; flex-wrap: wrap; }
	174	+.doc-head h1 { font-size: 32px; letter-spacing: -0.025em; margin: 0; font-weight: 760; }
	175	+.doc-head h1::first-letter { text-transform: uppercase; }
	176	+.doc-head-meta { color: var(--muted); font-size: 13px; margin: 6px 0 0; font-family: ui-monospace, Menlo, monospace; }
	177	+.doc-layout { display: block; margin-top: 18px; }
	178	+.doc-layout.has-toc { display: grid; grid-template-columns: 250px minmax(0, 1fr); gap: 36px; align-items: start; }
	179	+.doc-layout .doc { margin: 0; max-width: 820px; }
	180	+.toc {
	181	+ position: sticky; top: 80px; background: var(--surface); border: 1px solid var(--grid);
	182	+ border-radius: var(--radius); padding: 16px 18px; max-height: calc(100vh - 104px);
	183	+ overflow-y: auto; box-shadow: var(--shadow-1);
	184	+}
	185	+.toc-title { font-size: 11px; text-transform: uppercase; letter-spacing: 0.08em; color: var(--muted); font-weight: 700; }
	186	+.toc ul { list-style: none; margin: 10px 0 0; padding: 0; }
	187	+.toc li { padding: 3px 0; line-height: 1.35; }
	188	+.toc li a { color: var(--ink-2); font-size: 13px; display: block; border-left: 2px solid transparent; padding-left: 8px; margin-left: -8px; }
	189	+.toc li a:hover { color: var(--accent-dark); text-decoration: none; border-left-color: var(--accent); }
	190	+.toc-d3 { padding-left: 14px !important; }
	191	+.toc-d3 a { font-size: 12.5px !important; color: var(--muted) !important; }
	192	+.hanchor { opacity: 0; margin-left: 8px; font-weight: 400; color: var(--muted); font-size: 0.85em; }
	193	+h2:hover .hanchor, h3:hover .hanchor, h4:hover .hanchor { opacity: 1; text-decoration: none; }
	194	+
	195	+/* markdown body */
137	196	.md { overflow-wrap: break-word; }
138	197	.md h1 { font-size: 28px; letter-spacing: -0.02em; }
139		−.md h2 { font-size: 21px; margin-top: 36px; border-bottom: 1px solid var(--grid); padding-bottom: 6px; }
	198	+.md h2 { font-size: 21px; margin-top: 38px; border-bottom: 1px solid var(--grid); padding-bottom: 7px; letter-spacing: -0.01em; }
140	199	.md h3 { font-size: 17px; margin-top: 28px; }
141		−.md code { background: #f1f0ec; border-radius: 4px; padding: 1px 5px; font-size: 0.88em; }
	200	+.md code { background: #f1f0ec; border-radius: 5px; padding: 1px 6px; font-size: 0.88em; }
142	201	.md pre code, .codeblock code { background: none; padding: 0; font-size: 13px; }
143	202	.codeblock, .md pre {
144		− background: #f6f5f2; border: 1px solid var(--grid); border-radius: 8px;
145		− padding: 14px 16px; overflow-x: auto; line-height: 1.5;
	203	+ background: #f6f5f2; border: 1px solid var(--grid); border-radius: 10px;
	204	+ padding: 15px 18px; overflow-x: auto; line-height: 1.55;
146	205	}
147	206	.md table { border-collapse: collapse; width: 100%; font-size: 14px; display: block; overflow-x: auto; }
148		−.md th, .md td { border: 1px solid var(--grid); padding: 6px 10px; text-align: left; vertical-align: top; }
	207	+.md th, .md td { border: 1px solid var(--grid); padding: 7px 11px; text-align: left; vertical-align: top; }
149	208	.md th { background: var(--page); }
150	209	.md blockquote { border-left: 3px solid var(--accent); margin-left: 0; padding-left: 16px; color: var(--ink-2); }
151	210
152		−/* log */
153		−.log-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(280px, 1fr)); gap: 20px; }
154		−.log-entry h3 { font-size: 15px; margin: 0 0 8px; }
155		−.log-entry .md { font-size: 13.5px; color: var(--ink-2); }
	211	+/* ---------------------------------------------------------------- log */
	212	+.log-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(290px, 1fr)); gap: 20px; }
	213	+.log-entry h3 { font-size: 15px; margin: 2px 0 8px; }
	214	+.timeline { margin: 26px 0; max-width: 880px; }
	215	+.tl-entry { display: grid; grid-template-columns: 26px minmax(0, 1fr); gap: 14px; }
	216	+.tl-rail { position: relative; }
	217	+.tl-rail::before { content: ""; position: absolute; left: 11px; top: 4px; bottom: -18px; width: 2px; background: var(--grid); }
	218	+.tl-entry:last-child .tl-rail::before { bottom: auto; height: 24px; }
	219	+.tl-dot {
	220	+ position: absolute; left: 5px; top: 6px; width: 14px; height: 14px; border-radius: 50%;
	221	+ background: var(--accent); border: 3px solid var(--page); box-shadow: 0 0 0 1px var(--grid);
	222	+}
	223	+.tl-card {
	224	+ background: var(--surface); border: 1px solid var(--grid); border-radius: var(--radius);
	225	+ padding: 18px 22px; margin-bottom: 18px; min-width: 0; box-shadow: var(--shadow-1);
	226	+}
	227	+.tl-time { font-size: 12px; color: var(--muted); font-family: ui-monospace, Menlo, monospace; }
	228	+.tl-card h3 { margin: 4px 0 10px; font-size: 16.5px; letter-spacing: -0.01em; }
	229	+.md-compact { font-size: 14px; color: var(--ink-2); }
	230	+.md-compact ul { padding-left: 18px; margin: 6px 0; }
	231	+.md-compact li { margin: 3px 0; }
	232	+.md-compact p { margin: 6px 0; }
	233	+.md-compact strong { color: var(--ink); }
	234	+.md-compact pre { font-size: 12px; padding: 10px 12px; }
156	235
157		−/* results table / kv table / file lists */
158		−.results-table, .kv-table { border-collapse: collapse; width: 100%; font-size: 14.5px; background: var(--surface); }
	236	+/* ---------------------------------------------------------------- tables/files */
	237	+.results-table, .kv-table {
	238	+ border-collapse: collapse; width: 100%; font-size: 14.5px; background: var(--surface);
	239	+ border-radius: var(--radius); overflow: hidden; box-shadow: var(--shadow-1);
	240	+}
159	241	.results-table th, .results-table td, .kv-table th, .kv-table td {
160		− border: 1px solid var(--grid); padding: 8px 12px; text-align: left;
	242	+ border: 1px solid var(--grid); padding: 9px 13px; text-align: left;
161	243	}
162	244	.results-table th { background: var(--page); }
163		−.kv-table th { background: var(--page); width: 180px; }
	245	+.kv-table th { background: var(--page); width: 190px; }
164	246	.file-list { list-style: none; padding: 0; }
165		−.file-list li { padding: 6px 0; border-bottom: 1px solid var(--grid); }
	247	+.file-list li { padding: 7px 0; border-bottom: 1px solid var(--grid); }
166	248	.tree { list-style: none; padding-left: 0; }
167	249	.tree ul { list-style: none; padding-left: 20px; }
168	250	.tree-root { margin-bottom: 14px; }
169	251	.tree li { padding: 2px 0; font-size: 14.5px; }
170	252	.mono-small { font-size: 12.5px; color: var(--muted); }
171	253	.file-card { padding: 0; overflow: hidden; }
172		−.file-head { display: flex; justify-content: space-between; padding: 10px 16px; border-bottom: 1px solid var(--grid); background: var(--page); }
	254	+.file-head { display: flex; justify-content: space-between; padding: 10px 18px; border-bottom: 1px solid var(--grid); background: var(--page); }
173	255	.file-card .codeblock { border: 0; border-radius: 0; margin: 0; }
174	256
175		−/* chart */
	257	+/* ---------------------------------------------------------------- chart */
176	258	.chart-fig { margin: 18px 0 8px; position: relative; }
177	259	.chart-fig svg { width: 100%; height: auto; }
178	260	.chart-fig .grid { stroke: var(--grid); stroke-width: 1; }
		`@@ -184,112 +266,98 @@ code, pre, .mono-small { font-family: ui-monospace, SFMono-Regular, Menlo, monos`
184	266	.chart-fig figcaption { font-size: 12.5px; color: var(--muted); margin-top: 4px; }
185	267	.chart-tip {
186	268	position: absolute; pointer-events: none; background: var(--ink); color: #fff;
187		− font-size: 12.5px; padding: 5px 10px; border-radius: 6px; white-space: nowrap; z-index: 5;
	269	+ font-size: 12.5px; padding: 5px 10px; border-radius: 7px; white-space: nowrap; z-index: 5;
188	270	}
189	271
190		−/* hljs light theme (subset, GitHub-like) */
191		−.hljs { color: #24292e; }
192		−.hljs-keyword, .hljs-meta .hljs-keyword { color: #d73a49; }
193		−.hljs-string, .hljs-attr { color: #032f62; }
194		−.hljs-comment, .hljs-quote { color: #6a737d; font-style: italic; }
195		−.hljs-number, .hljs-literal { color: #005cc5; }
196		−.hljs-title, .hljs-function .hljs-title { color: #6f42c1; }
197		−.hljs-built_in, .hljs-type { color: #e36209; }
198		−.hljs-section { color: #005cc5; font-weight: 600; }
199		−.hljs-params { color: #24292e; }
200		−.hljs-variable, .hljs-template-variable { color: #e36209; }
201		−
202		−/* footer */
203		−.site-footer {
204		− border-top: 1px solid var(--grid); margin-top: 48px; padding: 20px 0 32px;
205		− background: var(--surface); font-size: 13px; color: var(--ink-2);
	272	+/* ---------------------------------------------------------------- comments */
	273	+.comment-grid { display: grid; grid-template-columns: minmax(280px, 420px) minmax(0, 1fr); gap: 24px; margin: 24px 0; align-items: start; }
	274	+.comment-form label { display: block; font-size: 13.5px; font-weight: 650; color: var(--ink-2); margin: 14px 0 4px; }
	275	+.comment-form input, .comment-form textarea {
	276	+ width: 100%; border: 1px solid var(--grid); border-radius: 9px; background: var(--page);
	277	+ padding: 10px 12px; font-size: 14.5px; font-family: inherit; color: var(--ink);
	278	+ transition: border-color 0.15s, box-shadow 0.15s;
206	279	}
207		−.site-footer .wrap { display: flex; justify-content: space-between; flex-wrap: wrap; gap: 8px; }
208		−.footer-meta { color: var(--muted); }
209		−
210		−/* ---- professional document view: header, TOC, anchors -------------------- */
211		−.doc-head { margin: 28px 0 6px; }
212		−.doc-head-row { display: flex; align-items: center; gap: 12px; flex-wrap: wrap; }
213		−.doc-head h1 { font-size: 30px; letter-spacing: -0.02em; margin: 0; text-transform: none; }
214		−.doc-head h1::first-letter { text-transform: uppercase; }
215		−.doc-head-meta { color: var(--muted); font-size: 13px; margin: 6px 0 0; font-family: ui-monospace, Menlo, monospace; }
216		−.doc-layout { display: block; margin-top: 18px; }
217		−.doc-layout.has-toc { display: grid; grid-template-columns: 240px minmax(0, 1fr); gap: 32px; align-items: start; }
218		−.doc-layout .doc { margin: 0; max-width: 820px; }
219		−.toc {
220		− position: sticky; top: 76px; background: var(--surface); border: 1px solid var(--grid);
221		− border-radius: 10px; padding: 14px 16px; max-height: calc(100vh - 100px); overflow-y: auto;
	280	+.comment-form input:focus, .comment-form textarea:focus {
	281	+ outline: none; border-color: var(--accent); box-shadow: 0 0 0 3px rgba(42,120,214,0.15);
222	282	}
223		−.toc-title { font-size: 11px; text-transform: uppercase; letter-spacing: 0.07em; color: var(--muted); font-weight: 600; }
224		−.toc ul { list-style: none; margin: 8px 0 0; padding: 0; }
225		−.toc li { padding: 3px 0; line-height: 1.35; }
226		−.toc li a { color: var(--ink-2); font-size: 13px; display: block; }
227		−.toc li a:hover { color: var(--accent-dark); text-decoration: none; }
228		−.toc-d3 { padding-left: 14px !important; }
229		−.toc-d3 a { font-size: 12.5px !important; color: var(--muted) !important; }
230		−.hanchor { opacity: 0; margin-left: 8px; font-weight: 400; color: var(--muted); font-size: 0.85em; }
231		−h2:hover .hanchor, h3:hover .hanchor, h4:hover .hanchor { opacity: 1; text-decoration: none; }
232		−html { scroll-behavior: smooth; scroll-padding-top: 72px; }
233		−.card-head-row { display: flex; justify-content: space-between; align-items: baseline; gap: 12px; flex-wrap: wrap; }
234		−
235		−/* ---- research-log timeline ----------------------------------------------- */
236		−.timeline { margin: 24px 0; max-width: 860px; }
237		−.tl-entry { display: grid; grid-template-columns: 24px minmax(0, 1fr); gap: 14px; }
238		−.tl-rail { position: relative; }
239		−.tl-rail::before {
240		− content: ""; position: absolute; left: 11px; top: 4px; bottom: -18px; width: 2px; background: var(--grid);
	283	+.comment-form .btn { margin-top: 16px; }
	284	+.comment-form .hp { position: absolute; left: -9999px; height: 0; width: 0; border: 0; padding: 0; }
	285	+.comment {
	286	+ background: var(--surface); border: 1px solid var(--grid); border-radius: var(--radius);
	287	+ padding: 16px 20px; margin-bottom: 14px; box-shadow: var(--shadow-1);
241	288	}
242		−.tl-entry:last-child .tl-rail::before { bottom: auto; height: 24px; }
243		−.tl-dot {
244		− position: absolute; left: 5px; top: 6px; width: 14px; height: 14px; border-radius: 50%;
245		− background: var(--accent); border: 3px solid var(--page); box-shadow: 0 0 0 1px var(--grid);
246		−}
247		−.tl-card {
248		− background: var(--surface); border: 1px solid var(--grid); border-radius: 10px;
249		− padding: 16px 20px; margin-bottom: 18px; min-width: 0;
	289	+.comment-head { display: flex; justify-content: space-between; gap: 10px; align-items: baseline; flex-wrap: wrap; }
	290	+.comment-name { font-weight: 700; font-size: 14.5px; }
	291	+.comment-time { font-size: 12px; color: var(--muted); font-family: ui-monospace, Menlo, monospace; }
	292	+.comment-body { margin: 8px 0 0; font-size: 14.5px; color: var(--ink-2); white-space: pre-wrap; }
	293	+.flash {
	294	+ background: #eef7ee; border: 1px solid #cfe8cf; color: var(--good-text);
	295	+ border-radius: 9px; padding: 10px 16px; margin: 14px 0; font-size: 14px; font-weight: 600;
250	296	}
251		−.tl-time { font-size: 12px; color: var(--muted); font-family: ui-monospace, Menlo, monospace; }
252		−.tl-card h3 { margin: 4px 0 10px; font-size: 16.5px; letter-spacing: -0.01em; }
253		−.md-compact { font-size: 14px; color: var(--ink-2); }
254		−.md-compact ul { padding-left: 18px; margin: 6px 0; }
255		−.md-compact li { margin: 3px 0; }
256		−.md-compact p { margin: 6px 0; }
257		−.md-compact strong { color: var(--ink); }
258		−.md-compact pre { font-size: 12px; padding: 10px 12px; }
	297	+.flash-err { background: #fdf0ef; border-color: #f2cfcc; color: #a13430; }
259	298
260		−/* ---- smartphone adaptation ------------------------------------------------ */
	299	+/* ---------------------------------------------------------------- footer */
	300	+.site-footer { border-top: 1px solid var(--grid); margin-top: 56px; background: var(--surface); font-size: 14px; color: var(--ink-2); }
	301	+.footer-grid { display: grid; grid-template-columns: 2fr 1fr 1fr; gap: 28px; padding: 32px 24px 20px; }
	302	+.footer-brand { font-weight: 750; color: var(--ink); font-size: 15px; }
	303	+.footer-tag { font-size: 13px; color: var(--muted); max-width: 340px; margin: 6px 0 0; }
	304	+.footer-h { display: block; font-size: 11px; text-transform: uppercase; letter-spacing: 0.08em; color: var(--muted); font-weight: 700; margin-bottom: 8px; }
	305	+.footer-col { display: flex; flex-direction: column; gap: 5px; font-size: 13.5px; }
	306	+.footer-meta { color: var(--muted); font-size: 12px; }
	307	+.footer-legal { border-top: 1px solid var(--grid); padding: 14px 24px 26px; font-size: 12.5px; color: var(--muted); }
	308	+
	309	+/* ---------------------------------------------------------------- mobile */
	310	+@media (max-width: 880px) {
	311	+ .split { grid-template-columns: 1fr; }
	312	+ .hero h1 { font-size: 32px; }
	313	+ .comment-grid { grid-template-columns: 1fr; }
	314	+}
261	315	@media (max-width: 720px) {
262	316	body { font-size: 15px; }
263	317	.wrap { padding: 0 16px; }
264		− .header-row { height: auto; flex-direction: column; align-items: flex-start; gap: 2px; padding: 10px 0; }
265		− .site-nav { width: 100%; overflow-x: auto; -webkit-overflow-scrolling: touch; flex-wrap: nowrap; padding-bottom: 4px; }
266		− .site-nav a { white-space: nowrap; padding: 6px 10px; font-size: 13.5px; }
267		− .site-header { position: static; }
268		− html { scroll-padding-top: 12px; }
269		− .hero { padding: 28px 0 4px; }
270		− .hero h1 { font-size: 26px; }
	318	+ .header-row { height: 56px; }
	319	+ .nav-toggle {
	320	+ display: flex; flex-direction: column; justify-content: center; gap: 5px;
	321	+ background: none; border: 0; padding: 10px; cursor: pointer; z-index: 60;
	322	+ }
	323	+ .nav-toggle span { display: block; width: 22px; height: 2px; background: var(--ink); border-radius: 2px; transition: transform 0.22s, opacity 0.18s; }
	324	+ .nav-toggle.open span:nth-child(1) { transform: translateY(7px) rotate(45deg); }
	325	+ .nav-toggle.open span:nth-child(2) { opacity: 0; }
	326	+ .nav-toggle.open span:nth-child(3) { transform: translateY(-7px) rotate(-45deg); }
	327	+ .site-nav {
	328	+ position: fixed; top: 56px; left: 0; right: 0; z-index: 50;
	329	+ background: var(--surface); border-bottom: 1px solid var(--grid);
	330	+ flex-direction: column; gap: 0; padding: 8px 12px 14px;
	331	+ box-shadow: var(--shadow-2);
	332	+ display: none;
	333	+ }
	334	+ .site-nav.open { display: flex; }
	335	+ .site-nav a { padding: 13px 12px; font-size: 16px; border-radius: 8px; border-bottom: 1px solid var(--grid); }
	336	+ .site-nav a:last-child { border-bottom: 0; }
	337	+ html { scroll-padding-top: 64px; }
	338	+ .hero { padding: 32px 16px 8px; margin: 0 -16px; }
	339	+ .hero h1 { font-size: 27px; }
271	340	.lede { font-size: 16px; }
272	341	.lede-eq code { font-size: 11px; white-space: normal; word-break: break-word; }
273		− .tiles { grid-template-columns: repeat(2, 1fr); gap: 10px; margin: 20px 0; }
274		− .tile { padding: 14px 16px; }
275		− .tile-value { font-size: 26px; }
276		− .card { padding: 16px 16px; }
277		− .page-title { font-size: 24px; margin-top: 24px; }
278		− .doc-head h1 { font-size: 22px; }
	342	+ .tiles { grid-template-columns: repeat(2, 1fr); gap: 10px; margin: 22px 0; }
	343	+ .tile { padding: 15px 16px; }
	344	+ .tile-value { font-size: 27px; }
	345	+ .card { padding: 18px 18px; }
	346	+ .page-title { font-size: 25px; margin-top: 26px; }
	347	+ .doc-head h1 { font-size: 23px; }
279	348	.doc-layout.has-toc { display: block; }
280	349	.toc { position: static; max-height: none; margin-bottom: 16px; }
281	350	.tl-entry { grid-template-columns: 18px minmax(0, 1fr); gap: 10px; }
282		− .tl-card { padding: 12px 14px; }
	351	+ .tl-card { padding: 13px 15px; }
283	352	.md h1 { font-size: 22px; } .md h2 { font-size: 18px; } .md h3 { font-size: 16px; }
284	353	.md pre, .codeblock { font-size: 12px; padding: 10px 12px; }
285		− .results-table, .kv-table { font-size: 13px; display: block; overflow-x: auto; }
	354	+ .results-table, .kv-table { font-size: 13px; display: block; overflow-x: auto; box-shadow: none; }
286	355	.kv-table th { width: auto; }
287	356	.file-head { flex-direction: column; gap: 2px; }
288		− .site-footer .wrap { flex-direction: column; }
289		− .chart-fig svg { min-width: 560px; }
	357	+ .footer-grid { grid-template-columns: 1fr; gap: 18px; }
290	358	.chart-fig { overflow-x: auto; -webkit-overflow-scrolling: touch; }
	359	+ .chart-fig svg { min-width: 560px; }
291	360	}
292	361	@media (max-width: 420px) {
293		− .tiles { grid-template-columns: 1fr 1fr; }
294		− .hero h1 { font-size: 23px; }
	362	+ .hero h1 { font-size: 24px; }
295	363	}

modified web/server.js +68 −0

		`@@ -11,6 +11,7 @@`
11	11	// ============================================================================
12	12	"use strict";
13	13
	14	+const fs = require("fs");
14	15	const path = require("path");
15	16	const express = require("express");
16	17	const C = require("./lib/content");
		`@@ -20,6 +21,26 @@ const app = express();`
20	21	const PORT = process.env.PORT \|\| 8120;
21	22
22	23	app.use("/static", express.static(path.join(__dirname, "public"), { maxAge: "1h" }));
	24	+app.use(express.urlencoded({ extended: false, limit: "16kb" }));
	25	+
	26	+// ---- comments storage (data/ is excluded from deploy rsync — persistent) ----
	27	+const DATA_DIR = path.join(__dirname, "data");
	28	+const COMMENTS_FILE = path.join(DATA_DIR, "comments.json");
	29	+fs.mkdirSync(DATA_DIR, { recursive: true });
	30	+
	31	+function readComments() {
	32	+ try {
	33	+ return JSON.parse(fs.readFileSync(COMMENTS_FILE, "utf8"));
	34	+ } catch {
	35	+ return [];
	36	+ }
	37	+}
	38	+function saveComment(entry) {
	39	+ const all = readComments();
	40	+ all.push(entry);
	41	+ fs.writeFileSync(COMMENTS_FILE, JSON.stringify(all, null, 2));
	42	+}
	43	+const lastPostByIp = new Map();
23	44
24	45	function page(res, opts) {
25	46	res.send(R.layout({ ...opts, buildInfo: C.buildInfo() }));
		`@@ -323,6 +344,53 @@ why an approach died, not just what survived.</p>`
323	344	page(res, { title: "About", active: "About", body });
324	345	});
325	346
	347	+// ---------------------------------------------------------------- comments
	348	+app.get("/comments", (req, res) => {
	349	+ const comments = readComments().slice().reverse();
	350	+ const list = comments.length
	351	+ ? comments.map((c) => `<article class="comment">
	352	+ <div class="comment-head"><span class="comment-name">${R.esc(c.name)}</span>
	353	+ <time class="comment-time">${R.esc((c.ts \|\| "").slice(0, 16).replace("T", " "))} UTC</time></div>
	354	+ <p class="comment-body">${R.esc(c.message)}</p>
	355	+ </article>`).join("")
	356	+ : `<p class="lede-small">No comments yet — be the first to leave one.</p>`;
	357	+ const posted = req.query.posted ? `<div class="flash">Thanks — your comment is posted.</div>` : "";
	358	+ const err = req.query.err ? `<div class="flash flash-err">${R.esc(String(req.query.err))}</div>` : "";
	359	+ const body = `<h1 class="page-title">Comments & discussion</h1>
	360	+<p class="lede-small">Questions, critiques, pointers to prior art, replication reports — all welcome.
	361	+Comments are public and moderated after the fact; be civil, no links required.</p>
	362	+${posted}${err}
	363	+<div class="comment-grid">
	364	+ <form class="card comment-form" method="POST" action="/comments">
	365	+ <h2>Leave a comment</h2>
	366	+ <label>Name<input name="name" maxlength="60" required placeholder="Your name"></label>
	367	+ <label>Comment<textarea name="message" maxlength="2000" rows="5" required
	368	+ placeholder="Your comment — feedback, prior art we missed, questions…"></textarea></label>
	369	+ <input type="text" name="website" class="hp" tabindex="-1" autocomplete="off" aria-hidden="true">
	370	+ <button type="submit" class="btn">Post comment</button>
	371	+ <p class="mono-small">Stored on this server only. No tracking, no account.</p>
	372	+ </form>
	373	+ <div class="comment-list">
	374	+ <h2 class="section-title" style="margin-top:0">${comments.length} comment${comments.length === 1 ? "" : "s"}</h2>
	375	+ ${list}
	376	+ </div>
	377	+</div>`;
	378	+ page(res, { title: "Comments", active: "Comments", body });
	379	+});
	380	+
	381	+app.post("/comments", (req, res) => {
	382	+ const ip = req.headers["x-forwarded-for"] \|\| req.socket.remoteAddress \|\| "?";
	383	+ const name = String(req.body.name \|\| "").trim().slice(0, 60);
	384	+ const message = String(req.body.message \|\| "").trim().slice(0, 2000);
	385	+ if (String(req.body.website \|\| "").length) return res.redirect("/comments?posted=1"); // honeypot
	386	+ if (!name \|\| message.length < 3) return res.redirect("/comments?err=Name and comment are required.");
	387	+ const last = lastPostByIp.get(ip) \|\| 0;
	388	+ if (Date.now() - last < 30_000) return res.redirect("/comments?err=Please wait a moment between comments.");
	389	+ lastPostByIp.set(ip, Date.now());
	390	+ saveComment({ name, message, ts: new Date().toISOString() });
	391	+ res.redirect("/comments?posted=1");
	392	+});
	393	+
326	394	// ---------------------------------------------------------------- misc
327	395	app.get("/health", (req, res) => res.json({ ok: true, app: "localvm-web", author: "Simon-Pierre Boucher" }));
328	396
329	397