The config seams: three ADR-0019 SILENT fixes and one deletion

Date: 2026-10-01. Lane: wt/config-seams, off a2a5877. Findings from docs/reviews/doc-coverage-2026-10-01.md (§3.1–§3.3) and docs/reviews/dead-pub-audit-2026-10-01.md §2; each fix is one commit, gates inline. CPU only (the seed wave owns the GPU).

1. config/loader.rs — a present-but-broken preset is COUNTED (179bdf3)

Was: try_file discarded both the read error and the parse error (read_to_string(p).ok()? + parse_str(&s).ok()), so --preset small with a syntax error in ./configs/small.toml fell through to the repo copy and trained on a config nobody asked for, with no word on any stream. The explicit path failed loudly on the same input — the asymmetry was the finding.

Is: try_file returns Option<Result<DormouseConfig, String>>: absent candidate stays a silent normal miss of a search path; a present-but-unusable one is printed with its reason before the search moves on —

[config] preset candidate configs/small.toml unusable (toml parse: TOML parse
error at line 2, column 11); trying next

Why COUNTED, not LOUD (one line, per ADR-0019): the name search is a best-match by design and the dangerous case is precisely the one where another candidate loads — a terminal-Err variant can never fire there, so the warning must fire at the moment of the skip; the run stays correct AND the reader of the log can tell which file actually won.

red→green, on the real code path: the old test binary (dormouse_core-225a4a1058c91f40, base a2a5877) run from a cwd holding a broken configs/small.toml: presets_match_original_values passed in total silence. The fixed binary, same command, same file: the warn above (twice — the test loads small twice) and still passes. Permanent gate: config::loader::tests::broken_candidate_is_reported_absent_is_not (broken → Some(Err) naming the parse; absent → None).

2. config/override.rs — --set knows the four queue arms (68cb247)

Was: 36 of the schema’s 40+ fields were reachable from --set; use_situ, use_attnres, use_mhc, mhc_streams were reachable only from a preset TOML — every A/B of those arms started with a hand edit under configs/. The refusal was LOUD (unknown config key), so nothing trained differently from what was asked; the defect was the inconsistency.

Is: four arms in the match (three bools via parse_bool, one usize), mirroring use_mor/mor_k. Gate: config::r#override::tests::arm_flags_resolve_from_set — one assertion per key, from DormouseConfig::default().

3. param.rs — the non-CUDA bf16_compute drop is COUNTED (21557ba)

Was: with bf16_compute on and no cuda feature, the bf16 matmul op does not exist and the fp32 path ran — right answer, no counter, no line. ADR-0019 row 29 recorded it SILENT-by-cfg; doc-coverage §3.3 made it the fourth open SILENT outside that enumeration’s fixed set.

Is: bf16_compute_dropped_once() — one stderr line per process (the call site runs on every forward):

[param] bf16_compute requested on non-cuda: fp32 path (counted, once)

The forward’s doc comment no longer says “not fixed here”. Gate: param::tests::bf16_compute_on_cpu_is_the_fp32_answer_and_says_so — the arm answers fp32-equal (max |Δ| < 1e-5) on the CPU test backend, and with --nocapture the line prints exactly once across two forwards through the branch. The once-per-process line is a deliberate ceiling: a per-call counter in probe.rs would cost an N_ARMS bump for a state that cannot recur.

4. dormouse-data — train_eval_split deleted (this commit)

Dead per dead-pub-audit §2 (called by nothing; the trainer takes an explicit --eval file), advertised by the crate doc at lib.rs:16. Function and doc line deleted together; grep train_eval_split crates/ tools/ is 0. The audit doc carries a one-line disposition note; its record is untouched. collect_files stays (it has live callers: bin/anchors.rs, tests, the trainer).

Gates

• cargo test -p dormouse-core -p dormouse-data --lib: 89 + 15 passed, 0 failed (baseline before any edit: 86 + 15, so +3 = the three new gates).
• RUSTFLAGS="-D warnings" cargo doc --no-deps -p dormouse-core -p dormouse-data: Finished, 0 errors. Note the honest scope: this gate enforces rustc lints; RUSTDOCFLAGS="-D warnings" still fails on 41 pre-existing missing-documentation items (schema.rs 30, override.rs 5 — all at lines 6/8/24, the old pub items — config/mod.rs 3, validation.rs 1, lib.rs 1). None are from this lane; closing them is a docs-lane decision, not smuggled in here.
• python3 tools/check_doc_refs.py: the only new path it names is this file (resolved by its existence); the archive failures are pre-existing.