Lane: удаление неживых крейтов vendor/dormouse-fused — УЛИКИ (§1.4)

Measured on main @ db64923, 2026-10-01. No repo file touched — the precondition (dormouse-kda in main’s Cargo.toml) was unmet at every reading.

The two instruments

# 1. the crates a training build compiles (vendor/ is `exclude`d, so the ROOT lock is honest)
grep '^name = "burn-' Cargo.lock          # 13 vendored names

# 2. every path-dependency edge into a vendored crate, and its kind
grep -rn 'burn-<x>' --include=Cargo.toml . # outside the crate itself

The reachability result, per candidate

W = normal [dependencies] of a crate under crates/dormouse-* (in the product). d = [dev-dependencies]. b = a bench package. – = the burn-fused facade only, and the facade is referenced by zero manifests in this repo (fate table’s own finding).

crate	src	t/b	other	live caller OUTSIDE itself	fate row	verdict
`burn-attnres`	2730	72	—	`crates/dormouse-core/Cargo.toml:32`, W	REFERENCE	KEEP — FINDING, row is STALE
`burn-mhc`	1437	0	—	`crates/dormouse-core/Cargo.toml:27`, W	REFERENCE	KEEP — FINDING, row is STALE
`burn-rope`	1139	0	—	7 sites, all `benches/` + 1 example (`d`,`b`)	WIRE	KEEP
`burn-parcae`	317	0	—	0 (`–`)	WIRE	KEEP
`burn-swiglu`	187	0	—	1 (`benches/cpu_probe/src/main.rs:74`, `b`)	REFERENCE	KEEP
`burn-eggroll`	327	0	—	0 (`–`)	REFERENCE	KEEP — the survivor of the `es` pair
`burn-ptrn`	399	0	—	0 (`–`)	REFERENCE	KEEP — reason to stay is live (§ below)
`burn-es`	345	0	99	0 (`–`)	REFERENCE	DELETE
`burn-sct`	2737	458	244	0 non-dev (`d`+`b` only)	DELETE (blocked)	DELETE

Two of the brief’s seven candidates have a LIVE PRODUCT CALLER. burn-attnres and burn-mhc are normal [dependencies] of crates/dormouse-core and both are in the root Cargo.lock. burn-mhc has been A/B’d to a 3-seed tie (git log: “mHC is a 3-seed TIE — 6.413/6.313/6.370”). The fate table’s rows for both still read REFERENCE; they are stale, not wrong about reachability at the time. Fixing them is a separate, cheap commit.

The delete set, and the evidence for each line of it

`burn-sct` — 3439 lines, 16 files

grep -rn "burn-sct" --include=Cargo.toml . outside itself → 3 edges: vendor/dormouse-fused/Cargo.toml:11 (workspace member), crates/burn-spectral/Cargo.toml:33 ([dev-dependencies]), benches/Cargo.toml:20 (bench package), plus the generated facade.
grep -rn "burn_sct" --include=*.rs outside itself → 11 sites, all in crates/burn-spectral/examples/tsct_diag.rs and benches/src/main.rs:209. Zero in crates/, in any src/, in any config, in any test that gates a product path.
crates/dormouse-core/src/param.rs:6 imports burn_spectral::SpectralLinear, not burn_sct — the product’s TSCT is the burn-spectral one.
research/decisions/class-b-2026-10-01.md:513 recommends delete; the crate’s own burn-sct/README.md:3-9 says “Not in the dormouse build … Recommendation: DELETE”; docs/architecture/library-crate-fate.md:61 already carries the verdict.
COST ESTIMATE CORRECTION (a finding). Fate:159-172 and the decision sheet:504 both price this at “~10 lines”. Measured: 85 of 1771 lines of tsct_diag.rs carry sct_a/sct_b/burn_sct (56 are mechanical sct_x: None, fillers, 2 struct fields, 2 whole match arms at 22 lines each, 16 need hand edits, 9 sit inside the dropped arms), plus a 14-line block in benches/src/main.rs, plus gpu-gate.sh:74,79-81. ~10 is off by an order of magnitude, and the 16 hand sites include two &[str; N] marker tables whose arity changes (7→5, 10→8) and two #[test] fixture lists naming heads.3.sct_a.weight.

`burn-es` — 444 lines, 4 files

grep -rn "burn-es" --include=Cargo.toml . outside itself → 2: workspace member + facade. Zero code callers.
The crate declares itself a duplicate. burn-es/src/lib.rs:14-19: “NOTE: this crate keeps a convenience eggroll_mutate … the dedicated burn-eggroll crate implements the paper’s normalized form … Prefer burn-eggroll for real ES loops; the function here stays for quick experiments and its test.”
A live doc orders the deletion and its precondition is met. docs/research/2026-09-27-adopt-vs-port.md:588: “the source settles the σ convention, then delete this one.” The convention IS settled — burn-eggroll is the paper’s (σ/√r)·A·Bᵀ and survives; burn-es documents σ·A·Bᵀ and defers. Same conclusion at docs/architecture/library-crate-fate.md:75 (“They overlap each other; one should go when it starts”) and adopt-vs-port.md:388-390 (the disagreement is √r).
One citation needs straightening: crates/burn-spectral/src/lib.rs:50 attributes the ternary 0.7·mean dead zone to “(burn-es convention)”. The audit that found it (docs/reviews/spectral-stack-audit-2026-10-01.md:65, F13) already grades it (c) — no external reference. Deleting burn-es makes the attribution dangle, so the honest wording is “no external reference exists”, which is the same edit F13 asks for.

Why `burn-ptrn` is KEPT despite the brief listing it

library-crate-fate.md:184 gives a live file:line reason to stay: ADR-0013 deleted the Q-head the crate scores rollouts with, and AGENTS.md §3.6 item 3 says the selection rule must be re-specified first. Fate’s own rule is that a documented reason to stay is exactly the exception that beats “delete is the default”. adopt-vs-port.md:587 argues “~20 lines it would take to write are cheaper than maintaining 399” — that is an argument for when it is wanted, not a decision to delete. Two live docs in mild tension = a §1.7 defect to report, not a licence to delete. Cost of being wrong is asymmetric: deleting costs a re-port; keeping costs 399 dead lines.

Counter to the brief’s arithmetic

The brief says “13 in the build, 7 in none, + burn-sct”. Measured: 21 crates total (20 under crates/ + the facade), 13 in the build, 7 not in the build = {sct, eggroll, es, parcae, ptrn, rope, swiglu} — sct is inside the seven, not beside it. Of those seven, two (attnres, mhc) are W, wired into dormouse-core and so are not in the “not in the build” set at all; rope and parcae carry fate dispositions of WIRE; swiglu has a live bench caller. The honest delete set is 2 crates, not 8.

Line counter

crate	src	t/b	other	total	files
`burn-sct`	2737	458	244	3439	16
`burn-es`	345	0	99	444	4
total	3082	458	343	3883	20

Against the lane’s target of −2000: exceeded by 3883, 194 %. No test file is deleted — burn-es and burn-sct are the only two candidates that have them, and both go whole.

The gate, and one correction to it

cargo check --workspace --all-targets inside the library will fail on burn-fused-benches / cpu-probe / launch-probe for reasons that predate this lane (they need --features cuda). The gate to run is CI’s own command, fused-library.yml:62: the same check plus --exclude burn-fused-benches --exclude cpu-probe --exclude launch-probe. Both are run per batch, plus gen_facade.py --check and cargo check -p dormouse-train --features dormouse-train/cuda in the root.

Execution (2026-10-02, branch wt/cut2)

Landed as three code commits plus the docs/README.md commit:

f13ec4e delete(vendor): burn-sct — 3439 lines, 16 files.
24834c6 refactor(burn-spectral): drop the sct8/sct16 arms from tsct_diag — 127 lines removed (85 carrying sct_a/sct_b, of which 56 were x: None fillers; the two 22-line match arms; 16 hand sites including the two arity-changed marker tables, 7→5 FFN / 10→8 Muon+).
5a37f94 delete(vendor): burn-es — 444 lines, 4 files; facade regenerated, gen_facade.py --check green; the dangling (burn-es convention) attribution in burn-spectral/src/lib.rs:50 is now “no external reference exists” (same edit F13 in docs/reviews/spectral-stack-audit-2026-10-01.md asked for).
docs: library-crate-fate.md rows for both crates now read DELETED 2026-10-02; docs/papers/tsct.md carries a tombstone for §5; docs/glossary.md now names burn_spectral::SpectralLinear; and crates/dormouse-core/src/param.rs:1’s stale “via burn-sct” header is via burn-spectral. tools/check_doc_refs.py carries four KNOWN_DEAD entries so dated audits naming the deleted paths stay green.

Gates (CPU): library workspace check per CI’s command (cargo check --workspace --all-targets --exclude burn-fused-benches --exclude cpu-probe --exclude launch-probe) — green, 4m38s; cargo check -p dormouse-train --features dormouse-train/cuda — green (3m24s, one pre-existing unused-import warning in burn-spectral); python3 tools/check_doc_refs.py — 0 named paths missing; npm run check in docs-site (after npm run build) — check: ok, check-mermaid: ok. Library gate: 4m38s, warnings only.

Line count against target: 3439 + 444 = 3883 removed (CI reset plus tsct_diag/bench trims bring the tracked total to ~3885+).