spec(port): port specs/port/trios/crates/trios-trainer/src/bin/trios-train.t27 (Closes #7770)

[dmitrii-f-t27]

Summary

Port gHashTag/trios:crates/trios-trainer/src/bin/trios-train.rs (at 8b229e9e9489ee, 95 lines) as specs/port/trios/crates/trios-trainer/src/bin/trios-train.t27 — Roadmap Stage 6 (goal #4548), target t27c gen-rust. Closes #7770.

The original is the training CLI entry point: three flags with overrides, an INV-8 LR gate, a dry-run branch and a Gate-2 verdict.

Per the issue's rule ("Port the decision, not the plumbing"), the one listed function is ported under its name as a pure function (no transports — the clap parsing, the TOML load and the training run are the callers' plumbing) with 6 tests (34 runtime asserts, 0 vacuous, 100% function coverage):

• the four-flag CLI shape: --config required, --seed and --steps optional overrides, --dry-run a switch (original lines 20-36);
• the override rule: a present CLI value replaces the config's value, an absent flag leaves it (lines 48-56);
• the INV-8 LR gate: the LR must sit in the phi band [0.001, 0.01], inclusive on both ends — the error message names the bounds (lines 59-61);
• the dry-run shape: six printed fields (seed, steps, LR, d_model, n_layers, checkpoint interval), the run exiting before any training (lines 65-74);
• the three-band Gate-2 verdict on the best BPB: below 1.50 the VICTORY CANDIDATE, below 1.85 the above-target evidence, else the below-target evidence (lines 87-93);
• the flow order in main: the LR gate first, then the dry-run branch, then training (lines 38-95).

Verification (all run live from the worktree root, zig 0.16.0)

• t27c parse: OK. t27c spec-status: IMPLEMENTED. t27c typecheck: 0 errors, 0 warnings.
• t27c coverage: 14/14 tested (100%) — no transports in this port.
• t27c test-report: 6 tests / 6 pass / 0 FAIL / 0 BLOCKED, 0 vacuous passes — the flag table, the override rule both ways, the LR band at and beyond both boundaries, the dry-run shape with its early exit, the Gate-2 bands at the exact thresholds, and the flow order.
• t27c gen | grep -c 'not yet implemented': 0; gen | wc -l: 150 (> 12).
• grep -cE '^\s*(pub )?fn (main)\(': 1. grep -cE '^\s*test\s+(["|[A-Za-z_])': 6 (≥ 1).
• Boundary: git status --short — exactly this one file. The repo's lefthook pre-commit passed.
• dupe_scan: all 13 introduced names unique across 2046 specs (checked before the commit); the bare body scan adds no duplicate group (722 bodies / 217 groups on base, identical with this spec present).

Dmitrii Fedorov (@dmitrii-f-t27)

[dmitrii-f-t27]

@gHashTag Please merge this PR into master via the golden chain — my rights on gHashTag/t27 are read-only (403), so only you can merge it.

Verification of this head (f0217c9):

• Required CI trio green: validate, check-linked-issue, parse-ratchet; also Check L1 TRACEABILITY and Corpus ratchet green, duplicate-bodies green.
• The t27b-native-run artifact of this head (commit f0217c9, run 39046066604) row for specs/port/trios/crates/trios-trainer/src/bin/trios-train.t27: reference pass, t27b pass, tests 6, invariants 0, asserts 34, blockers [], detail empty, reference_disagree: [].
• The only red checks are the known base-red pair (cli-tri ci: repair the two workflows whose failure was in the workflow #3324 include_str!, t27b-native-ratchet registration debt) — both red on master without this PR; no other check is red.

After the merge I will: verify the spec file live on the merge SHA, confirm #7770 auto-closed, and file the 14-step report.

Dmitrii Fedorov (@dmitrii-f-t27)