Documentation round: correct the seed-grid contract in SPEC, and add convention 9

[emrefbulut]

Documentation only; no source file is touched. Four commits, one per file.

The find worth leading with

SPEC.md said, in the file that is supposed to be the authority:

Default operating point is split_seed=42 and train_seed=0.

That is the Phase 5 regression written down as the contract — the same sentence the README carried and that 0.5.0 exists to correct. The real default is 5 split seeds × 3 training seeds = 15 pairs.

--split-seeds and --train-seeds were not in SPEC at all, which is the wrong omission to make: they are the flags that decide sample size, and this release exists largely because that number changed once without anyone noticing.

Checked mechanically rather than by reading — every --flag on audit, build, measure-leakage and train is now mentioned somewhere in SPEC. That also turned up build's long form --output, documented only as -o.

README — two limitations the audit report names and the README did not

"Known limitations" is what someone reads before deciding whether this tool fits their recording. It listed two. The audit report names two more on every run:

• audit cannot tell you a measurement is possible, only that it is impossible. It trains nothing, so it rules out the trivial case and cannot find the SNR band where a model is partly right — the only band where inflation is visible. The report prints NOT CHECKED task difficulty; the README now says so first.
• Physical independence is outside what any of these checks can reach. Two recordings sharing no samples, no air time and no timestamp can still be near-duplicates. This is the limit that eliminated DASH7 ds_indoor — two 8-second captures 43 seconds apart, same antenna and position — after it passed every count-based test.

Also fixed: a roadmap line pointing at "the first known limitation below" when that section is above it.

ROADMAP — one item was listed twice

"Verification with an own hardware capture" (Now item 2) and "Hardware capture on one device" (Next) are the same work, and nothing said so, making the open list look one item longer than it is. The Next entry now cross-references it and keeps its own reason for existing.

CONTRIBUTING — convention 9, from this week

The eight conventions were each written after breaking one. The ninth is "a claim repeated is not a claim checked", and the example is the __original_version error corrected in #22: written once, restated in four more files, and five files agreeing looked like corroboration when it was one edit copied. One gh api call settled it.

The entry names the cheap sources this repo actually has — gh api for what a PR says, artifacts/ for what a number is, git log -S for when a line appeared — and asks that a correction state what the old version claimed rather than quietly overwriting it. The PR template listed eight conventions and now lists nine.

Checked and found correct — no commit

Not everything needed changing, and saying so is part of the report:

checked	result
README status block	0.5.0, "On PyPI" — verified against PyPI's API: ['0.1.0'…'0.5.0'], latest 0.5.0, tag v0.5.0 exists
README quickstart	audit, train, --group-by collection all present and correct (collection takes no colon; path:/csv: do)
README roadmap boxes	every ticked item verified, incl. GROUP_SCHEMES == ('path','csv','collection')
methodology §2/§3 numbers	every inflation row re-derived from artifacts/*.md — all four tables match
methodology section numbering	test_the_three_documents_agree_on_the_category_numbers passes
methodology sigmf section	never carried the __original_version or "two maintainers" claims
methodology open questions	current
dead relative links	none, across *.md, docs/, .github/
stale 0.4.0 references	four remain, each a legitimate historical reference
the regression sentence	grep for it across the repo returns nothing

Verification

ruff check .            All checks passed!
ruff format --check .   82 files already formatted
pytest -q               388 passed, 1 skipped

🤖 Generated with Claude Code