⚡ Measure test weights at f7a10637

[taras]

One commit, one file: test-weights.json, exactly as the Measure test
weights run produced it.

Why

The committed weights were provenanced to 084397e1, which predates the end of
Slice B of #854. Nothing measured after it — Slices B (#859), C (#860), D (#861)
and E (#864) each merged without a weights commit, and four measurement attempts
in between were cancelled. Three suites that landed in those slices had no
recorded weight at all.

That is not free. A file without a recorded weight is charged the heaviest
weight the current corpus recorded, so a new test is never treated as free and a
deleted outlier cannot inflate the fallback forever. With three unmeasured files
the charge was large and wrong in one direction:

runtime	charged each	three files charged	actually take	overstated by
deno	238,191 ms (cli-npm-bin.test.ts)	714,573 ms	31,312 ms	683 s
node	167,498 ms (plan-markdown.test.ts)	502,494 ms	18,748 ms	484 s
bun	103,624 ms (plan-markdown.test.ts)	310,872 ms	16,252 ms	295 s

The partition is longest-processing-time-first, so it was packing shards around
eleven minutes of Deno work that does not exist, and the files it displaced went
somewhere worse.

What changes

Before:

• source.commit 084397e1, run 36655497896. repl-agent-interface.test.ts,
  repl-agent-journey.test.ts and repl-forms.test.ts absent.

After:

• source.commit f7a10637, run
  36801798533.
  All three measured — under Deno, repl-agent-journey 16,939 ms,
  repl-agent-interface 9,505 ms, repl-forms 4,868 ms — and every other
  file's weight is this run's.

How it was produced

Dispatched against main at exactly f7a10637, the merged Slice E head, with
main CI already green on it. The workflow prepares the checkout in the usual
order, runs every applicable file alone under each runtime's own runner on
ubuntu-latest, and writes only after every measured file passed.

Provenance is supplied, never inferred: the commit, run URL, attempt, runner
label and the three runtime versions arrive in the environment, and a missing one
is a refusal rather than a default. The committed file is the uploaded artifact
byte for byte — verified with cmp against the download, before and after
the commit. No millisecond here was typed.

What must stay true

• Weights describe a machine CI runs on. Measured on ubuntu-latest through
  the workflow, never on a laptop — which is what the provenance block exists to
  make checkable.
• Shard counts are measured, not chosen. This measurement implies floors of
  14 (Deno), 8 (Node) and 5 (Bun); the installed 15 / 10 / 5 each
  satisfy their own, so nothing here forces a recalibration. Raising a count
  takes its own five consecutive runs on one fixed head, so ci.yml is
  untouched.

Scope

Included

• test-weights.json, unchanged from the artifact.

Intentionally unchanged

• ci.yml and the installed shard counts.
• Every source, test and specification file. This PR changes no behavior.

Risks and limitations

• Weights drift as the corpus does; this is a snapshot of f7a10637, not a
  standing guarantee. Remeasure after anything that moves the corpus, the
  exclusions or a runner command.
• Three already-measured files moved by more than four seconds under Deno
  (cli-npm-bin 238s → 272s, cli-help 20s → 29s, build-npm 39s → 43s).
  That is ordinary runner variance at this scale and is why the partition is
  recomputed from the whole file rather than patched per entry.

Scope confirmation

• Every changed file supports the purpose described above.
• Unrelated cleanup and formatting changes are excluded.
• Generated or mechanical changes are clearly identified — the whole diff is
  one generated artifact.
• The description matches the final diff and test results.