feat(bench): commit the session profiler every measurement came from

[lloydsk]

Why

Every number in docs/dogfooding.md was produced by two throwaway Python profilers in /tmp. A benchmark whose measuring instrument is uncommitted and unversioned is not reproducible, and the method section of that doc now tells readers to re-derive the tables — this is the tool it points at.

What

bun scripts/measure-sessions.ts <dir-or-glob…> [--csv] — one row per session plus medians/ranges per arm, dependency-free Bun/TypeScript.

Per session: label (from the directory), inferred arm (bgrun when the transcript hands off to a job, else vanilla), executions (foreground suite runs and bgrun handoffs, reported separately), blocked-on-output, wall, context characters, per-call (wait, size, command), and whether the failure diagnostic reached context (DIAGNOSTIC_MARKER_UPSTREAM / _DOWNSTREAM) — for a red fixture, that is whether the detail arrived at all.

Two definitional notes, both deliberate:

• idle_s is wall − blocked (time in the session spent not waiting on the suite), while the profilers' reading — agent-chosen sleeps/waits — is kept separately as sleep_s, so the old numbers stay cross-checkable instead of quietly changing meaning.
• executions counts execution, not mention: cat long-job.sh is not a run, and a bun test that only appears in prose is not either.

Verification

• Cross-checked field-by-field against the Python profiler over 17 sessions: 0 mismatches on span, calls, executions, blocked-on-output and context chars; sleep_s equals the profiler's agent-chosen idle on all 17.
• bun test scripts/measure-sessions.test.ts — 10 tests, synthetic transcripts for a blocking run, a multi-call session, a no-execution session, and bgrun detection.
• tsc --noEmit clean.

Note on CI

package.json's test script becomes bun test extension/index.test.ts scripts/measure-sessions.test.ts, so the instrument is covered by the same job as the extension. Parsing is what turns transcripts into the doc's numbers, so it gets tests; a silent parser regression would rewrite the benchmark.

[github-actions]

CI report

Check	Result
tsc --noEmit	success
tests	success
npm pack --dry-run	success (7 files in tarball)

Ref: 2c68c5b4d282ae8daea1984884a33cf7c06bd394

[lloydsk]

Superseded: the fixture generator, the measurement instrument and the write-up are now one branch, bench/dogfooding (7 commits). No content is lost — this PR described one third of a deliverable.