Record the DASH7 SNR sweep as run once and unusable, and add convention 10

[emrefbulut]

What changed

Four documents now say that the DASH7 SNR grid was run once and that the
result is unusable, instead of saying it has not been run. No code changes.

• docs/leakage-real-snr.md — status block replacing "It has not been run"
• docs/methodology.md §7 — a new entry under its own heading
• CONTRIBUTING.md + .github/PULL_REQUEST_TEMPLATE.md — convention 10
• ROADMAP.md — Now item 1 no longer reads as never attempted

Why

The grid ran to completion in a scratch working tree. Its output was written
where the script was told to write it, was read, and was never committed; the
tree was cleaned up afterwards. Nothing survives — no commit on any branch, no
other worktree, no stash, no reachable object — so there is no table, no
per-run JSON and no log, and therefore no environment stamp, seed pairs or run
count to audit.

Two decisions are recorded with it:

• The figures are not transcribed from memory. A number whose source file
  does not exist is not reproducible from this repository, which is what
  docs/methodology.md promises in its first line. Writing them down would
  satisfy a reader and mislead them (conventions 3 and 9).
• The grid is not re-run. The plan is pre-registered: n, the SNR list and
  the stem were fixed before any run so the result could not choose its own
  stopping rule. Having read a result removes that property from the reader, so
  a repeat is not the run the plan describes (convention 5). The item stays
  open to someone who has not read the lost table.

The §7 entry is under a separate heading because it is a process failure, not
a code defect; mixing it into that list would blunt both. It carries the lesson
the loss actually teaches: an uncommitted result is worse than a result that
was never produced, because the work is believed to have been done. That is
convention 10.

How it was verified

Each claim of absence was checked rather than assumed:

git log --all --oneline -- artifacts/leakage_real_runs.json artifacts/leakage_real_table.md
# (no output -- never committed on any branch)

ls artifacts/ | grep real
# leakage_real_cliff_*, leakage_real_pilot_*, leakage_real_probe_*,
# leakage_real_stride_* only -- the two names the recipe promises are absent

git stash list            # the single stash does not contain them
git fsck --lost-found     # the dangling blobs are a 61-line synthetic stride file

Docs-only change, and the test suite reads these files, so:

uv run ruff format --check .   # 82 files already formatted
uv run ruff check .            # All checks passed!
uv run pytest tests/test_repo_hygiene.py tests/test_preflight.py -q
# 74 passed, 1 skipped in 349.57s

The skip is test_the_loraiq_skip_reason_names_the_variable_to_set, which
skips because IQFORGE_LORAIQ is configured on this machine; CI runs it
unconfigured. The full suite runs here in CI.

Conventions it touches

• 3 Do not claim what you did not measure — the lost figures are not
  written down, and only what the repository can still show is stated
• 5 Do not enlarge a sample after seeing the result — the reason the grid
  is not simply re-run
• 6 Report your own mistakes — the entry is a mistake made in this
  project, recorded rather than quietly dropped
• 9 A claim repeated is not a claim checked — "it has not been run" was
  true when written and had been restated since; each absence here was
  re-checked against git rather than copied forward
• 10 A measurement is committed the moment it exists — added by this PR,
  because this is the failure that taught it

🤖 Generated with Claude Code