Add resume benchmark harness and record baseline

[uncoooloj]

docs/Benchmarking.md defined standards but no numbers were ever recorded, so token-efficiency claims were unmeasurable.

• New scripts/benchmark-resume.py (stdlib-only, self-contained): measures bytes-that-enter-context per fixture under three strategies — raw full read, message-only projection (bounded tool summaries), keyword/evidence digest. Markdown table + --json output, tokens approximated as bytes/4.
• docs/Benchmarking.md gains 'Baseline measurements (2026-06-10)': on current fixtures, projection is 68.9% of raw and digest 30.9% overall; on noisy JSONL fixtures projection alone is 25–26% and the wrong-newest codex fixture drops to 12.6%. Fixtures are ~0.5–1.7KB while real transcripts run 1–70MB, so these ratios understate real-world reduction.

🤖 Generated with Claude Code

[hacktivist123]

Nice baseline — really helpful to have real numbers on record, and the honesty about tiny fixtures vs real 1–70MB transcripts is appreciated.

One thing worth tracking: benchmark-resume.py re-implements the projection/digest logic inline rather than importing from session-digest.py. That's a reasonable choice for a self-contained measurement harness, but the two implementations can drift over time — if the live digest strategy changes and the benchmark doesn't, the recorded ratios will stop reflecting real behavior.

Suggestion: either add a one-line note in the docstring / Benchmarking.md that the inline projection is a measurement copy and must stay in sync with session-digest.py, or add a small parity check that runs both on a fixture and asserts they agree within a tolerance. Not blocking — just flagging for future maintainers.