docs: the M15 milestone record — the field fixes, v2.0.1 and the first blog post (#321)

[radiusred-wordy]

Closes #321.

The M15 milestone record: docs/milestones/15-v2-0-1-what-the-fleet-migration-taught.md, written under the doc-synthesizer contract from the milestone's own trail — #305's body and comments, the seven delivery task issues and their pull requests with every review submission, and the release artefacts read back through the git, releases and actions APIs rather than quoted from the evidence comment.

What it records. Seven requirements as opened, with no edit to the milestone body at all — its userContentEdits query returns a total count of zero. Seven delivery tasks, six here and one in radiusred/codecrew-www, in merge order with their rounds. Twenty-five Decision records and six Deviation records, grouped by the question each task had to settle. Fourteen review submissions — four change requests, ten approvals submitted, seven standing and three dismissed by a rebase. One QA round, in which every requirement came back satisfied. The release, from the tag object through to the operator's own verification. The blog post, and the three Decisions about what a public page may say. Eight adopted backlog captures and four the milestone filed and left open. Eight protocol-discipline observations, and a "What the record does not contain" section with eight gaps in the trail.

Counts. Every one is re-derived from the source rather than carried across, which is the standard PR #304's two rounds set on the M14 record. Records are counted by tracker.ExtractRecords' own rule, which in this milestone reads one paragraph recording no deviation as a Deviation record, and one paragraph that reads like a Deviation label as none. Both are named in the record and in a Decision on the task: a count nobody can reproduce from the source is worse than one that needs explaining.

Also in this pull request. The M15 row appended to ROADMAP.md, already Done. A ### The M15 record section at the top of the fresh ## [Unreleased] in CHANGELOG.md, ending (#321). And the doc-synthesizer refresh: two clauses added to docs/introduction.md's verb summary — task finish names any issue its merge would take besides the task, and status names a task's holder from its start record and reports the repository's stale task branches. Nothing on either page was false, and the record says what was run to establish that: M15-R6's own flip had already moved both version claims to v2.0.1, the installed extension prints v2.0.1 (protocol 2.0), and SPEC §10's catalogue still counts forty-three refusal codes because this milestone added none. README.md is unchanged, and the Decision on the task says why.

Checks. gh codecrew milestone evidence 15 from the installed v2.0.1, run against this branch's milestone: seven requirements counted, every cited link reachable. Every github.com URL in the record — ninety of them — was opened through the API by hand as well; every relative link and heading anchor into the sibling milestone records checked to exist. gofmt -l ., go vet ./... and go test ./... clean, as the boundary check on a docs-only change. One commit, conventional, lowercase after the type, 87 characters, (#321), authored as radiusred-wordy[bot].

Decisions recorded on the task — three, and all three are mine to make: the scope of the introduction and README refresh, with what was run to justify it; attributing the reviewer seat's harness only for the four rounds the trail names, and declaring the other ten unattributed rather than inferring a harness from how a review reads; and counting the two record-grammar edge cases the way the scanner counts them rather than the way the prose reads. No deviations from the plan, and no ask-the-human point was reached.

For the reviewer. This dispatch stops at the open pull request: the review and gh codecrew task finish follow from the coordination layer. The two claims most worth a second opinion are the counts — every one is stated in the record's opening paragraph and each is re-derivable — and the harness attribution, where the record deliberately leaves ten of the fourteen rounds unattributed.

[radiusred-checky]

Changes requested. The document is the best-sourced record this project has produced — I re-derived every count in its opening paragraph from the API and all but one of them came back exact — but three claims are wrong on the branch as it stands, and one of them is wrong in a way that only got worse while the PR sat open. Reviewed as the reviewer seat (radiusred-checky), a distinct App from the author (radiusred-wordy).

Finding 1 (blocking) — the record's counts are already false at merge time, and the gap it names has since been filled

Two records landed on the milestone's trail after the record's commit (889db9d, authored 02:06:04Z) and before this review:

• radiusred/gh-codecrew#306 comment 5564002686, 02:09:01Z, operator: **Decision (operator, after the merge):** the reviewer's non-blocking round-two note on PR #315 … is declined, not merely left.
• radiusred/gh-codecrew#305 comment 5564026934, 02:11:36Z, operator: **Deviation:** the same harness failure again on the record PR #322's first round …

Both parse as records under tracker.ExtractRecords' own rule, which is the rule the record says it counts by. Six statements in docs/milestones/15-v2-0-1-what-the-fleet-migration-taught.md are now false:

1. Opening paragraph: "its three comments, carrying one Deviation record, the operator's verification of the release, and one QA comment" — #305 now has four comments and two Deviation records.
2. Opening paragraph: "the twenty-five Decision records and six Deviation records across the milestone issue and those task issues" — now twenty-six and seven.
3. Merge-order list, the #306 / PR #315 bullet: "Four Decisions" — now five.
4. ## Decisions, first paragraph: "twenty-two on the six task issues in this hub" — now twenty-three. And in the same paragraph: "None carries a parenthetical qualifier: every record in this milestone is a bare **Decision:** or **Deviation:** label, which is not true of M14's, where two qualified labels were invisible to a plain search." **Decision (operator, after the merge):** is precisely that qualified form, so the sentence now says the opposite of the truth about its own milestone.
5. ## Deviations, first line: "Six Deviation records: one by the operator on #305, three on #309 and one each on #310 and #311" — now seven, two on #305.
6. ## What the record does not contain, third bullet: "It is still there. The review left it explicitly to the author — 'take it or leave it, and either is defensible' — and no Decision records the leaving. A blessing and a deferral are both decisions, and this is recorded as neither." Comment 5564002686 is exactly that Decision: it declines the third passage, gives the reason (the **What it writes.** bullet describes what the 2.0.0 binary wrote, which is true of 2.0.0, where the two marked passages are the adopter's steps and had to change), a Trade-off: and a Rejected:. The gap this bullet is about has been closed, and the bullet is the one place in the document that would have to say so.

The standard here is the house one, not mine: M14's record opens "#269's eight comments", and #269 carries nine today — but the ninth is milestone close 14's own closing comment at 23:34:56Z, after PR #304 merged at 23:32:50Z. Nothing landed on M14's trail between synthesis and merge, so that record was true when it merged. This one would not be.

The fix is small and is the implementer's call in shape: either re-derive the six lines above against the trail as it stands and fold the operator's #306 Decision into ### A migrated spoke, and a released changelog section edited to describe it and the third What the record does not contain bullet; or keep the snapshot and say so explicitly, naming the two comments and their ids as landing after synthesis. What it cannot do is ship the counts as written. Note that the second one is a Deviation about this PR's own review round, so a record that counts it is recording its own review — the same self-reference M14's record navigated by leaving its own task's rounds out; either treatment is defensible if it is stated.

Finding 2 — capture #316's provenance runs backwards

radiusred/gh-codecrew#316 was created at 2026-09-07T00:06:27Z. PR #313's only review (5127113957) was submitted at 00:10:17Z — three minutes and fifty seconds later. Two sentences assert the reverse:

• ## Protocol-discipline observations region, ### An App is not an assignee… — "The reviewer confirmed the reading in round one — '…' — and #316 is that capture." This one is fine as written: it claims confirmation, not causation.
• The captures section: "#316 … comes from the implementer's third Decision on #307 and the reviewer's confirmation of it". The capture cannot come from a review that did not exist yet.
• ## The review rounds, the PR #313 bullet: "and a third is the reading that became capture #316" — same inversion.

The record gets the other two right and it is worth keeping them right by contrast: #318 (review round one 00:17:38Z → capture 00:18:47Z) and #319 (round two 00:33:43Z → capture 00:34:24Z) both follow their reviews. The remedy is one clause each: #316 came from the Decision on #307, and the review independently reached the same reading four minutes later.

Finding 3 — the blog post has five sections, not six

## The release, and the blog post: "in six sections that follow the adopted capture's shape: why a protocol major, what changed, what an adopter does, what deliberately did not change, and what the fleet migration taught, ending on the seven captures v2.0.1 answers."

docs/blog/posts/2026-09-07-protocol-2-0-and-the-migration.md in radiusred/codecrew-www carries five headings and no others: ## Why a protocol major, ## What changed, ## What an adopter does, ## What deliberately did not change, ## What the migration taught. The clause about the seven captures is the closing part of the fifth, not a sixth. The record's own list names five. Either "five sections" or "an unheaded opening and five sections", whichever the author meant.

What I re-derived, and what held

Every count in the opening paragraph except those in Finding 1 checks out against the API, and I ran the sources rather than reading them.

• Seven requirements, unedited. userContentEdits on #305 returns totalCount: 0; the body carries M15-R1…R7 and the Gates section is the scaffold placeholder, verbatim as quoted.
• Seven delivery tasks plus #321. #305's subIssues totals 8: #306–#311 CLOSED, radiusred/codecrew-www#27 CLOSED, #321 OPEN.
• Seven merged PRs, twenty-four commits. Per-PR commit counts from the API are 3/2/4/6/6/2 for #313/#312/#315/#314/#317/#320 and 1 for codecrew-www#28 — 24, and each matches the per-task figure in the merge-order list. Merge SHAs 8f9d9e9, 55a04ee, 9e225d9, 7610047, 9c322e0, 2dc147a, 9931d40 and every merge timestamp match to the second.
• Fourteen review submissions. 1/2/2/3/3/2/1 across the seven PRs, all by radiusred-checky[bot]: four CHANGES_REQUESTED, ten approvals submitted, seven standing and three DISMISSED. The three dismissal events are at 00:13:31Z, 00:31:42Z and 00:42:12Z, giving standing times of 2m13s, 2m00s and 8m29s — the record's figures exactly. Every review id it links resolves to the round it names.
• Force-pushes. 0 on #313, 1 on #312, 2 on #315, 3 on #314, 3 on #317, 0 on #320 — as stated. The four CHANGELOG quotations are verbatim and attributed to the right rounds.
• Timings. #305 opened 23:51:44Z; last delivery merged 01:25:44Z — one hour thirty-four minutes to the second. First task start (#308) 23:59:09Z. The five parallel PRs opened 00:03:50Z–00:09:02Z, 5m12s apart. #311's Deviation 00:56:23Z → the operator's Decision 00:57:29Z, one minute six seconds.
• The release, read from git/releases/actions. Tag ref → tag object fabdc17eba781a98af792e5542d730fab0285be9, message v2.0.1 (protocol 2.0), tagger radiusred-cody[bot] at 01:08:33Z, pointing at 2dc147ac7fa5539e7f1b171f72aa8d984840b2a2. Run 34072047784 (.github/workflows/release.yml) created 01:08:40Z on that SHA, conclusion success. Release gh-codecrew 2.0.1 published 01:09:55Z, neither draft nor prerelease, five assets with exactly the names given.
• QA cells, verbatim against comment 5563791131. All seven verdicts are satisfied and the table's status column carries the word unqualified. Every quotation in the notes column and in ## QA: one round, seven verdicts is verbatim: "Main build/test/vet/gofmt are clean", "highest issue comment count is #119 with 69", "the paging tests drive 101 comments and 150-row listings so dropping --paginate truncates", "backlog captures #316, #318 and #319 remain open and outside the adopted set", "apart from recorded #306 correction notes", "refuses CLOSED and prints dry run: nothing written", the R7 sentence about codecrew-www#29.
• Captures. All eight adopted (#287, #264, #301, #296, #297, #295, #303, codecrew-www#26) are closed and carry exactly two comments each. All four filed-and-open (#316, #318, #319, codecrew-www#29) are open and were filed by the operator. All seven adopted hub/spoke captures were created 2026-09-06 between 11:24:17Z and 22:30:41Z, with #301 and #303 after v2.0.0 published at 18:47:47Z.
• Gates. Neither repository has an issue labelled cc:needs-decision (0 and 0). All seven delivery-task Ask-the-human sections resolve to none, and the six quoted fragments are verbatim; PR #312's and PR #315's bodies each say "No deviations from the plan".
• Documented commands, executed. gh codecrew milestone evidence 15 built from this branch prints, byte for byte, the two lines the record quotes: requirements counted: M15-R1, … (7) and all 20 cited links resolve across 9 issues — evidence is reachable. milestone close 15 --dry-run reproduces the quoted gate block, differing only in #321 (in review) where the record shows (ready) — the same before-the-PR capture M10–M14's records all carry, so not a finding.
• Links. All 112 links in the document resolve: 90 github.com URLs fetched through the API, https://codecrew.works and the live post 200, and all 21 relative links into the sibling records exist with their heading anchors present.
• Claims about the code. SPEC §10's table still has 43 rows. --paginate is on Comments, SubIssues, listIssues and OpenPRsForBranch; the two deliberate single-page reads each carry the comment the record describes, and TestMilestoneIssuesPaginateAndRecentIssuesDoesNot is the test holding one of them there.

Record and mechanics

• Plan before the first commit. #321's body was edited with the plan at 01:48:08Z by radiusred-wordy; the sole commit is authored 02:06:04Z.
• Decisions as comments, not only in the body. Three on #321 (5563874552, 5563874676, 5563874789) at 01:49:38–40Z, matching the three the PR body claims.
• Closing references. closingIssuesReferences lists #321 alone. The body's only close-family word is the leading Closes #321, so capture #303's defect and the stricter dispatch rule are both satisfied.
• Commit. One, conventional, lowercase after the type, 87 characters, referencing (#321), authored as radiusred-wordy[bot].
• CHANGELOG. ### The M15 record at the top of the fresh ## [Unreleased], ending (#321).
• ROADMAP. The M15 row is appended after M14's, already Done, and its document link matches the file added.
• docs/introduction.md. The refresh is contract-mandated, not optional — the doc-synthesizer contract requires the introduction's verb claims to be true at every milestone boundary, and the old sentence enumerated three things task finish does where SPEC §6 now has four and said nothing about status's holder line or its stale-branch report. Both new clauses read true against SPEC §6's status and task finish rows. README.md untouched, with the reasoning recorded.
• Boundary check on a docs-only change. gofmt -l . clean, go vet ./... clean, go test ./... all green.

One nit, not a finding

"the sharpest statement of what a dry run is for in this codebase" (### A migrated spoke…) is a superlative reaching past the milestone with nothing cited behind it, where every other comparative in the document links the record it compares against. Scoping it to this milestone, as its neighbours are, would keep the document's own rule intact. Take it or leave it.

Findings 2 and 3 are one clause each. Finding 1 is the one that has to be answered before this merges: the record's whole claim on a reader is that its counts are re-derivable from the trail, and two of them stopped being so at 02:09:01Z.

[radiusred-wordy]

Round one answered at bc12d1f, one commit on top of 889db9d. Finding 1 is taken the first way the review offered: the counts are re-derived rather than snapshotted, and the record now names the instant it counted them.

Finding 1 — re-derived at 2026-09-07T02:26:50Z, against a trail the coordination layer has frozen. The two records the review names are both folded in, and a third is mine: a **Deviation:** on #321 recording that the numbers moved because the trail moved, which is counted in with the rest. Every affected line: #305 now four comments and two Deviation records; the milestone totals are twenty-nine Decision records and eight Deviation records across the milestone issue and all eight task issues, of which twenty-six and seven are on the milestone issue and the seven delivery tasks — the shape M14's figure has, given beside the whole so the self-reference the review flagged is stated rather than decided in silence; the #306 / PR #315 bullet and the M15-R1 table cell now read five Decisions; the ## Decisions opening reads twenty-three on the six hub task issues, names both operator Decisions, and replaces the parenthetical-qualifier sentence with its opposite — one record carries a qualifier, **Decision (operator, after the merge):** on #306, invisible to a search for the bare label exactly as M14's two were; the ## Deviations opening reads eight, two of them on #305.
The operator's #306 Decision is folded into ### A migrated spoke, and a released changelog section edited to describe it as its fifth Decision, with its Trade-off and Rejected. The second #305 Deviation has its own entry beside the first. And the third What the record does not contain bullet no longer says no Decision exists: it says the gap was closed three minutes after this document's first commit, cites the Decision, gives its reasoning, and says the bullet stays because the paragraph it replaces asserted the opposite. A ninth protocol-discipline observation records the general fact — a record's counts have a shelf life, and this one's expired between commit and review, where M14's trail gained nothing between synthesis and merge.

Finding 2 — #316's provenance no longer runs backwards. The captures paragraph now says the capture was "filed at 00:06:27Z from the seat's own report and four minutes before the review that independently reached the same reading", and the PR #313 bullet says the review's third finding "reaches the same reading as capture #316, which the operator had already filed four minutes earlier from the implementer's own report". The observation sentence claimed confirmation rather than causation and is unchanged.

Finding 3 — five sections. Now "an unheaded opening and five sections", and the fifth is named as the one ending on the seven captures.

Nit — taken. "the sharpest statement of what a dry run is for in this codebase" is now "this milestone's clearest statement of what a dry run is for", scoped like its neighbours.

One thing I added beyond the findings. The milestone close 15 --dry-run block now carries a clause saying it is quoted as it read before this document's own pull request opened, and that the gate names #321 (in review) with the pull request open — the difference the review noted and did not raise.

Re-verified at the new head. gh codecrew milestone evidence 15 from the installed v2.0.1 still prints the two lines the record quotes, unchanged. All 93 github.com URLs in the record — three more than at round one, the new comment permalinks — were opened through the API; all relative links and heading anchors into the sibling records exist. gofmt -l ., go vet ./... and go test ./... clean. The new commit is conventional, lowercase after the type, 72 characters, (#321), authored as radiusred-wordy[bot]; no line of its body passes 100. A plain push: origin/main has not moved from 2dc147a, so no rebase.

[radiusred-checky]

Approved. Round two. Every count round one listed is now exact against the API as of the commit under review, the three other findings are answered, and gh codecrew milestone evidence 15 passes from this branch. Reviewed as the reviewer seat (radiusred-checky), a distinct App from the author (radiusred-wordy).

Finding 1 (blocking, round one) — resolved: the counts are re-derivable

I re-derived every figure round one named by replaying tracker.ExtractRecords' own rule (^\*\*(Decision|Deviation|Gate resolved)(\s*\(…\))?:\*\* at a paragraph start, on NormalizeLineEndings'd bodies) over the paginated comments of all nine M15 issues. The trail is unchanged since round one except the one **Deviation:** on #321 at 02:26:40Z, and nothing on any M15 issue is later than that — so the record's "counted at 2026-09-07T02:26:50Z, against a trail that is frozen from there" is true as of this review.

Per issue: #305 0/2, #306 5/0, #307 3/0, #308 3/0, #309 4/3, #310 4/1, #311 4/1, #321 3/1, codecrew-www#27 3/0 (Decisions/Deviations). Totals 29 Decisions, 8 Deviations; excluding #321, 26 and 7. Each of the six broken statements now checks out:

1. Opening paragraph — "its four comments, carrying two Deviation records, the operator's verification of the release, and one QA comment holding all seven verdicts": #305 returns exactly four comments — the two **Deviation:** records at 00:46:06Z and 02:11:36Z, the operator record for M15-R6 at 01:21:41Z, and radiusred-testy[bot]'s comment at 01:36:41Z carrying M15-R1…R7 all satisfied.
2. Opening paragraph — "twenty-nine Decision records and eight Deviation records … of which twenty-six and seven are on the milestone issue and the seven delivery tasks": exact, both figures.
3. Merge-order list, #306 / PR #315 — "Five Decisions, the fifth the operator's and recorded after the merge": five, and 5564002686 is at 02:09:01Z against PR #315's merge at 00:29:53Z.
4. ## Decisions — "twenty-three on the six task issues in this hub, three on the spoke's task, and three on this document's own — the milestone issue itself carries none": 5+3+3+4+4+4 = 23, 3, 3, 0. "Two of the twenty-nine are the operator's": davison wrote exactly two Decisions (5564002686 on #306, 5563544774 on #311); the other twenty-seven split radiusred-cody[bot] 21 / radiusred-wordy[bot] 6, as stated. And the parenthetical-qualifier claim is now the right way round and is the only one: **Decision (operator, after the merge):** on #306 is the sole non-bare label in the milestone — Decisions or Deviations — and the other twenty-eight Decisions are bare.
5. ## Deviations — "Eight Deviation records. Seven are on the milestone issue and the delivery tasks — two by the operator on #305, three on #309 and one each on #310 and #311, those five by the implementer seat — and the eighth is on this document's own task": exact, including the authorship split (davison ×2, radiusred-cody[bot] ×5, radiusred-wordy[bot] ×1) and "None is recorded on #306, #307, #308 or the spoke's task".
6. ## What the record does not contain, third bullet: rewritten to say the gap was closed after the list was written, and its paraphrase of 5564002686 is faithful — the quoted grounds ("describes what the 2.0.0 binary wrote", the adopter's steps, the **Trade-off:** and **Rejected:** clauses) are verbatim against the comment. The same Decision is now also folded into ### A migrated spoke… as a fifth Decision, quoted verbatim there too.

The self-reference is stated rather than silently resolved — both figures given, the instant named, and the **Deviation:** on #321 that records the recount counted in and cited. The ## Deviations and observations sections each carry it. That is the treatment round one said was defensible if stated.

I also re-checked the two claims the recount could have disturbed and did not: the **Deviation:** on #305 at 02:11:36Z is quoted verbatim (the 28 GiB clause and the withdrawal of the page-cache explanation included), and "the only harness attribution anywhere in the milestone" is now correctly narrowed to "the milestone's delivery".

Finding 2 (round one) — resolved: #316's provenance reads forward

radiusred/gh-codecrew#316 was created 00:06:27Z by the operator; PR #313's only review 5127113957 was submitted 00:10:17Z. Both sentences now run in that order: the captures section has "filed at 00:06:27Z from the seat's own report and four minutes before the review that independently reached the same reading", and the ## The review rounds PR #313 bullet has "a third reaches the same reading as capture #316, which the operator had already filed four minutes earlier from the implementer's own report". No remaining sentence claims the capture came from the review.

Finding 3 (round one) — resolved: five sections

docs/blog/posts/2026-09-07-protocol-2-0-and-the-migration.md in radiusred/codecrew-www carries exactly five headings — ## Why a protocol major, ## What changed, ## What an adopter does, ## What deliberately did not change, ## What the migration taught — preceded by an unheaded opening paragraph. The record now says "an unheaded opening and five sections", and its list of five matches the headings verbatim, including the heading's own wording ("what the migration taught", not "the fleet migration").

The nit (round one) — taken

"the sharpest statement of what a dry run is for in this codebase" is now "it is this milestone's clearest statement of what a dry run is for" — scoped to material the record itself enumerates, which is the document's own rule for comparatives.

What I ran

• gh codecrew milestone evidence 15, built from this branch (go build ./cmd/codecrew) and run against the live milestone: exit 0, printing requirements counted: M15-R1, … (7) and all 20 cited links resolve across 9 issues — evidence is reachable — byte for byte the block the record quotes.
• gh codecrew milestone close 15 --dry-run from this branch reproduces the quoted gate block with radiusred/gh-codecrew#321 (in review), which is exactly what the record's new sentence says it would show with this pull request open. The disclaimer is accurate.
• Boundary check, docs-only change: gofmt -l . clean, go vet ./... clean, go test ./... green across internal/cli, internal/config, internal/gh, internal/tracker.
• Mechanics. New commit bc12d1f is conventional, lowercase after the type, subject 72 characters, references (#321), and no body line exceeds 100. It touches one file — docs/milestones/15-v2-0-1-what-the-fleet-migration-taught.md, +141/−44 — so round one's checks on CHANGELOG.md (### The M15 record under a fresh ## [Unreleased], ending (#321)), ROADMAP.md and docs/introduction.md stand unchanged. closingIssuesReferences still lists #321 alone, and the body's only close-family word remains the leading Closes #321. The two relative links added both resolve, and the ## Protocol-discipline observations count went 8 → 9 to match the nine bullets that are actually there.

Two nits, neither a finding

• ## Deviations, line 807: "the operator's #306 Decision landed three minutes later and the second #305 Deviation five minutes after that". Read against the commit at 02:06:04Z both offsets are right; read with "that" as the #306 Decision the second is 2m35s. The observations bullet states the same thing with absolute timestamps and is unambiguous, so this is a wording preference only.
• Line 1218: "the review that found them opened on the observation that '…'". That sentence is the round-one review's closing line rather than its opening; the quotation itself is verbatim, and Finding 1 was the review's first and blocking finding, so the sense is right.

Neither blocks. The record is honest about its own moving target, and the counts in it are re-derivable by anyone who runs the same rule against the frozen trail.