Batch: follow-ons-from-417

[sv-tmueller]

Approved by the owner on 2026-10-01 (sign-off: dispatch). The proposal below is the contract, verbatim.

Packages

• Make each tm- workflow's meta a literal first statement #424 - Make each tm- workflow's meta a literal first statement (PR fix: make each workflow meta a literal first statement #426 merged, 9c6669d)
• Token report: price Sonnet 5.5 and use the lead's real output tokens #425 - Token report: price Sonnet 5.5 and use the lead's real output tokens (PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 merged, cd630c1)
• Auto-must-fix severity floor in review prompts #407 - Auto-must-fix severity floor in review prompts (PR feat: add severity floor to review prompts #428 merged, 9c5eb5c)

---

What I checked before slicing

• The workflow loader bug is real on main. In all three tm- workflows (tm-review-changes.js, tm-review-codebase.js, tm-map-codebase.js), export const meta is built from SPEC and TIER_MODELS, and const SPEC comes before it. The Workflow runtime needs meta to be a plain literal and the first statement in the file. The installed 2.9.0 plugin has the same code, so /tm-review-changes probably won't load as shipped. A/B arm: live ultracode on Opus 5.5 on a replayed merged issue #405 and feat: verify stage in tm-review-changes with spec-format item transforms #419 got around it with a modified copy. Nothing outside the runtime reads meta, so the fix stays contained.
• The token report's output figures. In a real lead transcript, output_tokens holds values in the hundreds to thousands. In a subagent transcript the values are 2 to 16, so they are placeholders. That backs up Decision 3 from batch Batch: measurement-tools #417. The price table has claude-sonnet-5 but not claude-sonnet-5-5.

Proposed batch: follow-ons-from-417 (3 packages, one wave)

1. Make each tm- workflow's meta a literal first statement

Scope. In each of the three workflow files, meta becomes a plain literal and the first statement. It holds the same content as today: name, description, and each phase's title, detail and model. A new test reads each file and checks three things:

• meta is the first statement.
• meta evaluates in an empty node:vm context, which proves it has no outside references.
• meta deep-equals what SPEC produces today, so editing SPEC without meta fails CI.

The architect picks whether the literal copies SPEC or SPEC reads its phases from meta. The sync test is required either way. After the tester passes, the lead runs the branch's tm-review-changes.js once, by path and unmodified, on the PR's own diff. The last such run cost about $2 and took about 3.5 minutes. Replying "dispatch" approves that cost.

Acceptance criteria

• In all three tm- workflow files, export const meta is the first statement and a plain literal, with no identifiers, calls or interpolation.
• A new test fails on today's main and passes after the fix. It checks position and evaluates the literal in an empty vm context.
• The test checks that each meta deep-equals the value SPEC produces (name, description, and each phase's title, detail and model).
• The lead's live run of the unmodified branch file loads. The PR body records that it loaded, plus cost and wall-clock time.
• npm test passes, and the plugin.json version is bumped.

Size: S. Track: full, because it touches .claude/workflows and executable code.

Non-goals

• Changing any stage, prompt or SPEC content.
• Live runs of tm-review-codebase and tm-map-codebase. Each fans out over the whole repo and costs much more. The refusal happens when the file loads, so the static test covers both.
• Porting the verify stage (review agent orchestration #344 P2).

2. Token report: price Sonnet 5.5 and use the lead's real output tokens

Scope. Sonnet 5.5 runs show up as "unpriced" and are left out of the total. #419 had to price them by hand. This package adds a claude-sonnet-5-5 entry to token-prices.json, read from the pricing page on the day of the change.

The report also estimates every agent's output as visible characters divided by 4. That leaves out thinking tokens and undercounts. The lead's transcript has the real usage.output_tokens, so the lead's rows use that number. Subagent transcripts only have placeholders, so their rows keep the estimate, and each row says whether its figure is measured or estimated.

Acceptance criteria

• token-prices.json has a claude-sonnet-5-5 entry with all five rates, taken from the source URL, and retrieved is set to the check date.
• A fixture test shows a Sonnet 5.5 record is priced and counted in the total.
• Lead rows use the transcript's output_tokens, subagent rows keep the characters-divided-by-4 estimate, and a fixture test covers both.
• The column header and notes say which figures are measured and which are estimated. The current line that says all output is estimated is replaced.
• npm test passes, and the version is bumped.

Size: S. Track: full, because token-report.mjs is executable code under .claude/skills.

Non-goals

• Re-pricing past batch reports.
• Changing how wall-clock time is worked out.
• Adding prices for models the policy doesn't pin.
• Any other way of recovering subagent thinking tokens.

3. #407 (already filed): auto-must-fix severity floor in review prompts

Scope. The issue body stays as filed, with two changes:

• I drop its merge-order note, because Verify stage in tm-review-changes, with spec-format item transforms #406 merged as feat: verify stage in tm-review-changes with spec-format item transforms #419.
• I add one point to it: the floor sets severity, not whether a finding is true. So the critic prompt says it can't downgrade a floor finding, and the new verify stage can still reject one on the facts (for example, the "deleted test" was never deleted). The verify prompt doesn't change.

Acceptance criteria: as filed, plus:

• The critic or consolidate prompt says a floor finding can't be downgraded.

Size: S. Track: full, because it touches .claude/agents and .claude/workflows.

Non-goals: as filed, plus no change to the verify prompt.

Merge note. All three packages bump plugin.json from 2.9.0, so the second and third merges each need a one-line version rebase. Packages 1 and 3 both edit the two review workflow files, but in different places (the top-of-file meta versus PROMPTS). Suggested order: 1, then 3, then 2.

Deferred, and why

• review agent orchestration #344 P2 (verify stage in tm-review-codebase): next batch. Its new Verify phase goes into the meta that package 1 rewrites, and its live-run criterion needs /tm-review-codebase to load. I'd also fold in the leftover finalizeReport guard-test nit from feat: verify stage in tm-review-changes with spec-format item transforms #419, since it's the same code path.
• The two trials (a second native-vs-kickoff pair, and /code-review versus the reviewer's quality pass): Record the headless trial traps from #400, #404 and #405 in tm-ab-test #418 set these up as the main work for this batch. I'm putting them after package 1 because the native-vs-kickoff judge is tm-review-changes, which currently loads only from a modified copy. That would add a known deviation to an expensive headless run. The /code-review trial also has no protocol yet. If you'd rather run them now with the workaround, say "revise".
• Hermes (Hermes adapter: swap spawn from delegate_task to the hermes subprocess #351 / PR feat: spawn Hermes seats via subprocess instead of delegate_task #354, Make SKILL.hermes.md a thin wrapper over the pipeline driver #355, and testing the verify output on a real Hermes host): still waiting on your provider credential and the GLM 5.3 model ID. Tell me when the credential works.

Things only you can do (not packages)

• Merge PR docs: use a placeholder provider URL in the hermes config example #353. It's a one-line change with CI green, and the real provider host is still on main.
• Merge PR docs: survey Claude-of-Tanks orchestration and plan the adoption #345, the adoption plan that Auto-must-fix severity floor in review prompts #407 and P2 cite (docs only).
• Close Desk study: Jev as a typed decision oracle for routing #370 or keep it. Batch Batch: measurement-tools #417 recommended closing it, because the Batch: lead-effort-and-ultracode #403 numbers already point to no-go on routing savings.

---

Decision log

Decisions made during the run are posted as comments on this issue.

[sv-tmueller]

Decision 1 (lead, run start 2026-10-01T10:52Z). The run uses the project-level role agents and skills from this checkout (main at 43cee0b, plugin 2.9.0), not the orchestrai: plugin copies. The installed plugin cache is also 2.9.0, so the two match. No dependency environment is built: CLAUDE.md marks install as N/A (zero-dependency npm test), so dispatches carry no Dependency environment: line. All three packages run concurrently (cap 3), full track.

[sv-tmueller]

Decision 2 (lead, #407 sub-plan). The architect's five open-point calls are accepted as written. Two are worth naming:

• docs/architecture/role-contracts.md gets the floor too. It is not in the signed-off surface list, but it is the host-neutral source that both reviewer.md bindings derive from. Leaving it out would ship a known-stale architecture doc, against the CLAUDE.md rule that a doc is corrected in the same PR. This adds the already-approved rule to its source doc. It is not a new policy and does not conflict with anything recorded there, so it stays in scope and is not parked. The human merge review should look at that hunk.
• The issue's claim that all five conditions are already in process-core.md is wrong for condition 1. "A weakened or deleted test" comes from the reviewer contract (role-contracts.md, "A weakened or deleted test is always a blocking finding"). The prompt text therefore does not cite process-core.md as the source. The existing, stricter "always blocking" sentence stays untouched.

The other calls: workers get a severity rule, not a detection mandate; "cannot downgrade" applies to severity only, so the critic can still dismiss a floor finding that is false, with a reason; tm-review-codebase gets one sentence saying to apply only the conditions the tree can show; the list wording is taken word for word from the issue. Version: 2.10.0 (minor), rebased above whatever is on main at merge.

[sv-tmueller]

Decision 3 (lead, #424 live run). The lead ran the branch's tm-review-changes.js by path, unmodified, in the lead checkout detached at the PR head (1afc284). HEAD was checked unchanged afterward and the checkout returned to main. It loaded on the first try; this is the acceptance criterion's proof that the fix works. Result: approve, 0 must-fix, 3 nits dismissed, $1.23 list price, 136,708 ms measured. Cost came from PR #427's branch copy of token-report.mjs (main's price table cannot price Sonnet 5.5), run read-only on a scratch dir holding only the run's 8 transcripts. Details are in the PR #426 body. While the checkout was detached, the #425 reviewer was dispatched with worktree isolation, which the full track does not normally use, so it could not touch the lead checkout.

[sv-tmueller]

Token report (batch #423)

The whole batch ran in one session, so this report covers all of it, from batch creation (2026-10-01T10:53:05Z) on. The output below is main's token-report.mjs, the canonical command. It shows Sonnet 5.5 as unpriced, which is exactly the gap #425 fixes. For a full total, see the rerun with PR #427's branch copy at the end.

Token report

• Session: b8a31d2a-581e-4d49-9574-81bb45613b99
• Window: from 2026-10-01T10:53:05Z
• Price source: https://platform.claude.com/docs/en/about-claude/pricing (retrieved 2026-09-27)

By role

Role	Calls	Input	Cache read	Cache write (5m)	Cache write (1h)	Output (est.)
lead	49	118	8,579,000	0	88,080	22,188
architect	66	132	3,711,974	271,525	0	10,905
developer	79	158	3,562,493	171,013	0	17,204
reviewer	57	114	2,355,248	172,467	0	5,512
tester	62	124	2,225,478	134,026	0	10,874
workflow:tm-review-changes	41	82	1,639,989	286,213	0	2,835

By model

Model	Calls	Input	Cache read	Cache write (5m)	Cache write (1h)	Output (est.)	Cost
claude-opus-5-5	181	382	15,056,847	503,530	88,080	39,821	$7.03
claude-sonnet-5-5 (unpriced)	173	346	7,017,335	531,714	0	29,697	-
Total	354	728	22,074,182	1,035,244	88,080	69,518	$7.03

By agent

Start (UTC)	Role	Model	Calls	Tool uses	Output (est.)	Wall-clock	Source
2026-10-01 10:53:05	lead	claude-opus-5-5	49	43	22,188	19m 40s (est.)	lead first-to-last span
2026-10-01 10:54:29	architect	claude-opus-5-5	21	26	3,070	4m 29s	task notification duration_ms
2026-10-01 10:54:29	architect	claude-opus-5-5	17	24	4,118	5m 2s	task notification duration_ms
2026-10-01 10:54:29	architect	claude-opus-5-5	28	43	3,717	6m 38s	task notification duration_ms
2026-10-01 10:59:34	developer	claude-sonnet-5-5	22	32	4,437	1m 51s	task notification duration_ms
2026-10-01 11:00:12	developer	claude-sonnet-5-5	24	28	6,585	2m 26s	task notification duration_ms
2026-10-01 11:02:05	developer	claude-sonnet-5-5	33	44	6,182	2m 49s	task notification duration_ms
2026-10-01 11:02:05	tester	claude-sonnet-5-5	15	22	3,597	1m 31s	task notification duration_ms
2026-10-01 11:02:52	tester	claude-sonnet-5-5	16	20	2,421	1m 29s	task notification duration_ms
2026-10-01 11:04:08	workflow:tm-review-changes (review:bugs)	claude-sonnet-5-5	7	7	302	40s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:security)	claude-sonnet-5-5	3	3	102	9s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:docs)	claude-sonnet-5-5	5	6	235	17s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:perf)	claude-sonnet-5-5	3	3	79	8s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:scope)	claude-sonnet-5-5	5	5	653	24s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:style)	claude-sonnet-5-5	4	4	110	12s (est.)	transcript span
2026-10-01 11:04:09	workflow:tm-review-changes (review:tests)	claude-sonnet-5-5	5	5	138	19s (est.)	transcript span
2026-10-01 11:04:47	reviewer	claude-opus-5-5	15	18	1,538	3m 1s	task notification duration_ms
2026-10-01 11:04:49	workflow:tm-review-changes (consolidate)	claude-opus-5-5	9	9	1,216	1m 34s (est.)	transcript span
2026-10-01 11:05:05	tester	claude-sonnet-5-5	31	49	4,856	2m 35s	task notification duration_ms
2026-10-01 11:07:29	reviewer	claude-opus-5-5	19	21	1,797	3m 28s	task notification duration_ms
2026-10-01 11:08:11	reviewer	claude-opus-5-5	23	32	2,177	3m 51s	task notification duration_ms

Limitations

• Output token counts are not in the transcripts; the "Output (est.)" column is estimated from visible text and tool-call input, divided by 4.
• The estimate excludes thinking tokens, so it undercounts real output token usage.
• Prices are Anthropic public list prices, not your actual billing (discounts, batch pricing and negotiated rates are not reflected).
• Wall-clock: "task notification duration_ms" is the time the harness measured for a dispatched seat. "transcript span" is the first-to-last timestamp span of each run in a transcript, used for workflow subagents and for a seat with no positive measured duration. "est." marks a span. The lead row spans its first to last line in the window, so it includes idle and waiting time. "n/a" means no source gave a positive value.
• Tool uses are counted from tool_use blocks in each transcript.
• Unpriced models (left out of the total): claude-sonnet-5-5.

Rerun with PR #427's branch script (read-only)

Same session and window, using token-report.mjs and token-prices.json from fix/425-token-report-pricing. With Sonnet 5.5 priced and the lead's output measured, the total is $10.54 at list price (Opus $7.51, Sonnet 5.5 $3.03). Subagent output is still estimated, so the figure undercounts.

By model

Model	Calls	Input	Cache read	Cache write (5m)	Cache write (1h)	Output (est. where marked)	Cost
claude-opus-5-5	181	382	15,056,847	503,530	88,080	63,726 (est.)	$7.51
claude-sonnet-5-5	173	346	7,017,335	531,714	0	29,697 (est.)	$3.03
Total	354	728	22,074,182	1,035,244	88,080	93,423 (est.)	$10.54

[sv-tmueller]

Batch report: follow-ons-from-417

All three packages reached a ready PR. None parked. No limit events and no run-long Sonnet fallback: every architect and reviewer verdict ran on Opus, so no Opus re-review is owed. No fix rounds were needed in any package.

Packages

• Make each tm- workflow's meta a literal first statement #424 literal workflow meta: PR fix: make each workflow meta a literal first statement #426 ready. meta is now a plain literal at byte 0 in all three tm- workflows. A new test checks its position, that it is a plain literal, and that it matches SPEC; it fails on main. The tester caught all 16 mutations. Live run: the runtime loaded the branch's tm-review-changes.js unmodified, by path, on the first try. That is the fix proven. The run approved with 0 must-fix, cost $1.23 at list price and took 136,708 ms (Decision 3). Reviewer APPROVE, both passes clean, 2 nits.
• Token report: price Sonnet 5.5 and use the lead's real output tokens #425 token report: PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 ready. It adds a claude-sonnet-5-5 price entry; the tester re-read the rates on the live page. The lead's output is now measured from usage.output_tokens, while subagent output stays estimated and is marked "(est.)". Snapshot diffs touch only Output, the notes, retrieved and Cost. Tester PASS. Reviewer APPROVE with one should-fix (below).
• Auto-must-fix severity floor in review prompts #407 severity floor: PR feat: add severity floor to review prompts #428 ready. The five-condition list is in every review prompt surface and in docs/architecture/role-contracts.md (Decision 2), with a drift test across all of them. The critic cannot downgrade a floor finding but can still dismiss a false one. PROMPTS.verify is unchanged. The tester's scratch merge with fix: make each workflow meta a literal first statement #426 conflicts only on plugin.json. Reviewer APPROVE with one should-fix (below) and 3 nits.

Decisions made during the run

1. The run used the project-level agents and skills from main (43cee0b, 2.9.0). No dependency environment, because install is N/A.
2. Auto-must-fix severity floor in review prompts #407: all five of the architect's calls accepted. role-contracts.md is in scope as the reviewer bindings' source doc. The issue's claim that process-core.md holds all five conditions is wrong for condition 1, so the prompts do not cite it.
3. Make each tm- workflow's meta a literal first statement #424 live run: the lead checkout was detached at the PR head for the run, and HEAD was checked unchanged afterward. Cost came from PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427's branch script on a scratch copy of the run's 8 transcripts. While the checkout was detached, the Token report: price Sonnet 5.5 and use the lead's real output tokens #425 reviewer ran with worktree isolation.

Should-fix before merge (recommended)

• PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427: no test catches bucket.outputEstimated ||= ... turning into =, because in every fixture the estimated record comes last in its bucket. Fix: an aggregate case with an estimated record before a measured one in the same bucket. Tester and reviewer both raised it.
• PR feat: add severity floor to review prompts #428: nothing pins the consolidate prompt's "may still dismiss it if it is false on the facts" escape. A scratch removal still passed npm test. Fix: a second sentinel, asserted in both consolidate templates (JS and JSON). It matters most in tm-review-codebase, which has no verify stage.

Nits are listed in each package's summary comment (#424, #425, #407).

Merge order and version bumps

All three bump plugin.json from 2.9.0. Suggested order:

1. fix: make each workflow meta a literal first statement #426 first (2.9.1). It fixes the loader the next batch depends on.
2. feat: add severity floor to review prompts #428 (2.10.0). It is already above 2.9.1, so the version conflict resolves by keeping 2.10.0. The workflow files auto-merge.
3. fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 last. Rebase it to 2.10.1.

Follow-on candidates for the next batch

1. review agent orchestration #344 P2 (verify stage in tm-review-codebase), with the finalizeReport guard-test nit from feat: verify stage in tm-review-changes with spec-format item transforms #419 folded in. Unblocked once fix: make each workflow meta a literal first statement #426 merges.
2. The two trials (a second native-vs-kickoff pair, and /code-review versus the reviewer's quality pass). Unblocked once fix: make each workflow meta a literal first statement #426 merges, and fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 too for cost figures. The /code-review trial needs a protocol first.
3. Hermes (Hermes adapter: swap spawn from delegate_task to the hermes subprocess #351 / PR feat: spawn Hermes seats via subprocess instead of delegate_task #354, Make SKILL.hermes.md a thin wrapper over the pipeline driver #355, and the verify {confirmed, note} output on a real host). Still waiting on the provider credential.
4. Small and optional: whether the runtime accepts a header comment above meta. The test requires byte 0, so this matters only if someone wants a header comment there.

Flags for the owner

• After merging fix: make each workflow meta a literal first statement #426, run /plugin update. The installed 2.9.0 plugin still has the computed meta, so /tm-review-changes, /tm-review-codebase and /tm-map-codebase probably stay broken until then.
• Still open from batch Batch: measurement-tools #417: merge PR docs: use a placeholder provider URL in the hermes config example #353 (the real provider host is still on main), merge PR docs: survey Claude-of-Tanks orchestration and plan the adoption #345 (the plan Auto-must-fix severity floor in review prompts #407 and P2 cite), and decide on closing Desk study: Jev as a typed decision oracle for routing #370.
• Cleanup: every worktree from this batch is removed, and the lead checkout is clean on main. The origin-safety rule kept two leftover worktrees from batch Batch: measurement-tools #417: .claude/worktrees/agent-a626d5b68e2009dc6 (69800c2, 4 commits not on origin) and .claude/worktrees/agent-ac44e4ae0c397f8f1 (93bb0b1, 5 commits not on origin). Both are clean. Batch Batch: measurement-tools #417 found 93bb0b1's tree identical to main's 43cee0b, so removing them by hand loses nothing. The local branches docs/352-placeholder-provider-url and feat/351-hermes-subprocess-spawn are kept; they back your draft PRs docs: use a placeholder provider URL in the hermes config example #353 and feat: spawn Hermes seats via subprocess instead of delegate_task #354.

Lessons (tester and reviewer)

• Lesson (Make each tm- workflow's meta a literal first statement #424, tester): a literal-shape guard needs a mutation test for each forbidden construct (calls, spreads, comments). Evaluating in an empty vm context alone lets .map() and Object.freeze() calls through.
• Lesson (Auto-must-fix severity floor in review prompts #407, tester): a sentinel-phrase assertion pins only the phrase. If a rule has a carve-out clause, pin the carve-out with its own assertion, or a matching edit to the JS and the JSON can remove it unnoticed.

What happened

• PR fix: make each workflow meta a literal first statement #426 ready - Make each tm- workflow's meta a literal first statement (Make each tm- workflow's meta a literal first statement #424)
• PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 ready - Token report: price Sonnet 5.5 and use the lead's real output tokens (Token report: price Sonnet 5.5 and use the lead's real output tokens #425)
• PR feat: add severity floor to review prompts #428 ready - Auto-must-fix severity floor in review prompts (Auto-must-fix severity floor in review prompts #407)
• Lesson (Auto-must-fix severity floor in review prompts #407, tester): a sentinel-phrase assertion pins only the phrase; pin a carve-out clause with its own assertion (1 more lesson above)
• Token report: Batch: follow-ons-from-417 #423 (comment) ($10.54 list price with Sonnet 5.5 priced, via PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427's script; main's script shows $7.03 with Sonnet unpriced)

Next steps

1. Review & merge: fix: make each workflow meta a literal first statement #426, feat: add severity floor to review prompts #428, fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427, in that order, with the version rebases above (merging closes each package issue via its Closes #N). Consider the should-fixes on fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 and feat: add severity floor to review prompts #428 first.
2. After fix: make each workflow meta a literal first statement #426 merges, run /plugin update so the installed tm- workflows load.
3. Close this batch issue by hand once everything is merged (it is the run record), or run /tm-advisor to confirm the merges, close it, and propose the next batch (review agent orchestration #344 P2 and the two trials first).

[sv-tmueller]

Closing: all packages merged

Owner request (2026-10-01): merge all three and close this batch.

• PR fix: make each workflow meta a literal first statement #426 merged as 9c6669d (closes Make each tm- workflow's meta a literal first statement #424), plugin 2.9.1. The squash message fixes the reviewer's nit 2 (the runtime refuses the file; it does not just "read it unreliably").
• PR feat: add severity floor to review prompts #428 merged as 9c5eb5c (closes Auto-must-fix severity floor in review prompts #407). The lead merged main into the branch and kept the version at 2.10.0 (4be5935); npm test 495 pass and the version guard passed locally, CI green before merge. The workflow files auto-merged with fix: make each workflow meta a literal first statement #426, as the tester's scratch merge predicted.
• PR fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427 merged as cd630c1 (closes Token report: price Sonnet 5.5 and use the lead's real output tokens #425). The lead merged main into the branch and resolved the version to 2.10.1 (1760be5); npm test 506 pass and the version guard passed locally, CI green before merge.
• main is at cd630c1, plugin 2.10.1, npm test 506 pass. No package parked, so no Part of batch lines needed stripping.

Not applied (merged as reviewed, at the owner's request): the two should-fixes from the report, plus the nits.

• fix: token report prices Sonnet 5.5 and uses the lead's real output tokens #427: a test with an estimated record before a measured one in the same bucket, so the "(est.)" marker cannot be lost silently.
• feat: add severity floor to review prompts #428: a second sentinel pinning the consolidate prompt's "may still dismiss it if it is false on the facts" escape.
• Nits: fix: make each workflow meta a literal first statement #426 nit 1 (the "display purposes" comment wording); feat: add severity floor to review prompts #428 nits 2 to 4.
  These are candidates for the next batch, alongside review agent orchestration #344 P2 and the two trials.

Owner step: run /plugin update so the installed tm- workflows pick up the literal meta. The installed 2.9.0 copy still has the computed one.

Cleanup: the lead checkout is on main at cd630c1, clean. The two scratch merge worktrees were removed after confirming their trees match the squash commits. The two batch #417 leftovers (agent-a626d5b68e2009dc6, agent-ac44e4ae0c397f8f1) are still kept by the origin-safety rule.