fix: token report prices Sonnet 5.5 and uses the lead's real output tokens

[sv-tmueller]

Closes #425

Part of batch #423.

The token report priced claude-sonnet-5-5 as unpriced and estimated the lead's output as visible chars / 4, though lead transcripts carry the real count.

• Price table: add claude-sonnet-5-5 ($2 / $2.50 / $4 / $0.20 / $10 per MTok). Re-read the pricing page on 2026-10-01: the existing claude-opus-5-5 and claude-sonnet-5 rates are unchanged; retrieved is now 2026-10-01.
• Parse: the lead's usage.output_tokens is read from the first line of each turn (every line repeats the final count). A lead turn with no count falls back to chars / 4 and is marked. Subagent output stays a chars / 4 estimate.
• Render: header is "Output (est. where marked)"; estimated cells (and any by-model or total figure that includes an estimate) carry " (est.)". Limitations text updated.
• Doc: one sentence in tm-ab-test/SKILL.md was stale and is corrected.
• Version bump 2.9.0 to 2.9.1 (patch, a fix).
• No new dependencies.

Checks: npm test passes (478 tests). Smoke run against a real recent lead session: lead row unmarked and in the thousands, claude-sonnet-5-5 priced, subagent rows marked (est.).

🤖 Generated with Claude Code

[sv-tmueller]

Tester report (round 1)

VERDICT: PASS
COMMIT: 46cc9ca
FINDINGS: none
UNTESTED CLAIMS:

1. A bucket stays marked "(est.)" when a measured record arrives after an estimated one. Mutating bucket.outputEstimated ||= record.outputEstimated to = in addToBucket leaves all 478 tests green. The mixed by-model test only has the estimated record last. The shipped code is correct. Repro: change that one line in a copy of the repo and run npm test.
2. Cosmetic: the edited tm-ab-test/SKILL.md line 242-244 runs past the wrapped width after "output, which it now reads from usage.output_tokens)." Only that sentence changed.
   Checked and clean:

• npm test gave 478 pass, 0 fail.
• node scripts/check-version-bump.mjs origin/main HEAD exited 0 (2.9.0 -> 2.9.1).
• Sonnet 5.5 rates match the live pricing page: 2 input, 2.50 5m write, 4 1h write, 0.20 cache read, 10 output. The Opus 5.5 rates (4/5/8/0.2/20) and Sonnet 5 rates are unchanged.
• The snapshot diff changes only Output headers and cells, the limitations notes, retrieved, and the lead-dedupe Cost cells. No Input, Cache or Wall-clock change, and no .jsonl changes.
• parse() passes isLead explicitly. A lead line with no output_tokens falls back to the estimate and is marked "(est.)". Total and by-model rows are marked when any part is estimated.
• The aggregate record() helpers now carry outputTokens and outputEstimated. 9 of 10 mutations were caught, including the per-role sum in both addToBucket and perAgent, the lead flag, the estimate fallback and the Total and by-model render. The one survivor is item 1.
• My smoke run matches the developer's figures: lead row 160,256 unmarked, claude-sonnet-5-5 priced, subagent rows "(est.)", total $29.14.

[sv-tmueller]

Reviewer report (round 1)

VERDICT: APPROVE
STAGE: quality
FINDINGS:

1. .claude/workflows/tests/token-report.test.mjs:234 (should-fix): no test catches it if bucket.outputEstimated ||= record.outputEstimated in addToBucket turns into =, so tester claim 1 holds. In every fixture, including both snapshots, an estimated record comes after the measured ones in its bucket, because parse() puts lead records first. A bucket that gets an estimated lead turn (no output_tokens) and then a measured one would lose its "(est.)" marker without any test failing. Fix: add an aggregate case where an estimated record comes before a measured record in the same bucket, and assert the bucket stays estimated.
   Evidence: token-report.mjs:215; in the records fixture (test.mjs:195-201), orphan-1, the estimated record, comes last in both the opus bucket and the totals.
2. .claude/skills/tm-ab-test/SKILL.md:244 (nit): the edited line is 112 characters. The lines around it wrap at about 76, so tester claim 2 holds. Fix: rewrap lines 242-245 to the file's width.
   Evidence: line 244 is "output, which it now reads from usage.output_tokens). Measured output was 3.76x the estimate on A/B: native plan-and-execute vs the kickoff pipeline on real issues #400 arm A".
   CHECKS: n/a