feat(m1): verify MILESTONE1.md matches the trinity.m1 code + offline workflow guard

[minion1227]

The gap

The Milestone-1 drop (feat(m1): …) added the miner-facing contract in docs/MILESTONE1.md — the < 1,000,000 param budget, 1/day rate, ≥ 0.02 win margin, the pack layout, and the 0.7·domain + 0.3·difficulty composite. Those same numbers live independently in trinity/m1/constants.py and the pack/metrics code, and nothing checks the two agree. A doc edit or a constant change could split them silently — and a miner who follows the document would then build a pack the validator rejects (or trains against the wrong composite).

What's here

src/trinity/m1/conformance.py parses MILESTONE1.md and asserts it against the code:

Check	Doc	Code
param_budget	< 1,000,000	constants.MAX_HEAD_PARAMS
rate_limit	1 … per day	RATE_LIMIT_MAX_SUBMISSIONS / RATE_LIMIT_WINDOW_DAYS
win_margin	≥ 0.02	constants.WIN_MARGIN
composite_weights	0.7·domain + 0.3·difficulty	constants.DOMAIN_WEIGHT
pack_layout_*	files under submissions/<miner>/m1/	what save_milestone1_pack writes

All pass on the shipped doc today — this is a guard, not a bug fix, and I'd rather say so than dress it up. Its value is prospective and it's cheap to keep: test_a_changed_param_budget_in_the_doc_is_caught, …_win_margin_…, …_composite_weight_… and …_phantom_pack_file_… show a real drift is caught rather than the guard being vacuous.

One documentation gap it surfaces

save_milestone1_pack can write W_k.npy (attentive packs with a key projection), but the doc's Pack-layout block lists only config.json / W_domain.npy / W_diff.npy / attention_query.npy. That's a file the code emits that the document omits. It's reported as a note, not a failure — an undocumented-but-real optional file is a doc gap, not a code defect. If you'd like the doc to list it, that's a one-line follow-up; this PR just makes the omission visible instead of letting it hide.

Also guards the offline workflow

Beyond the doc-parity checks, the tests exercise the pack → preflight path MILESTONE1.md prescribes (steps 2–3), which had no end-to-end coverage:

• a pack round-trips through disk, and the default 5-domain penultimate pack is the documented ≈ 10,240 params
• it clears every offline gate (run_m1_gates)
• the < 1,000,000 boundary is exact — 999,990 accepted, 1,000,000 rejected
• the ≥ 0.02 king/challenger decision matches decide_m1_winner (no-king, exactly-at-margin, a hair under)
• CODE_PACK_FILES cannot drift from the packer: a test saves a required-only pack and an attentive+key pack and asserts the file set on disk equals the declared list

Scope

Detection only. No M1 behaviour, constant, or doc is changed. m1/ and the milestone scripts are untouched by every other open PR.

Verification

• pytest tests/test_m1_conformance.py -q → 25 passed
• existing test_m1_gates / test_m1_leaderboard / test_m1_metrics → 13 passed, unchanged
• mypy on the new module → clean; the new files are ruff-clean
• import trinity.m1.conformance leaves torch out of sys.modules

Note on CI: main is currently ruff-red independent of this PR — src/trinity/m1/metrics.py:6 has an unused from typing import Iterable that shipped with the M1 commit, so ruff check src/ exits 1 before pytest runs. That is a one-line fix in a file this PR does not touch; every open PR is red on it until it lands. (Flagged separately.)

🤖 Generated with Claude Code