[darko-mijic]

A whole-repository review, its execution, and a second review of both

This pull request is the one record of three arcs on main b5e0a8be (PR #36):

1. A whole-repository review, first opened as PR #38: twelve lanes read the Specs, code, tests, docs, frozen delivery records and recent pull-request record. It produced an inventory of the durable content in plans/, reviews/ and the concept documents, a seven-change proposal, and six questions for the owner.
2. The execution of that review, first opened as PR #39: nine fix lanes carried out the confirmed findings and the owner's rulings.
3. A second review of both, and its fixes. A fresh cross-family review read PRs docs(articles): publish the product and projection narrative #34, review: the whole repository, and the durable-content inventory #38 and fix: execute the whole-repository review's findings #39 again. Most of what it confirmed was the first pass's defect classes reaching further than the first fixes; four fix rounds closed it.

The owner chose to review and merge the three as one pull request. PRs #38 and #39 close in favour of this one; their descriptions stay as those arcs' records. Both branches are wholly inside this diff, including this pass's edits to PR #38's evidence (684844eb...3dedfaa4). History stays as it happened: merge commits are kept, and nothing was squashed, rebased or amended, because git is the Protocol's event log.

PR #34 is not in this diff. The article series stays a separate pull request, and this pass's 39 findings against it were handled there: fixes, among them a rewrite of the article evidence checker and its tests, were pushed straight to its branch and reviewed lane by lane by the other model family, but through no pull request of their own. PR #34 is not a draft and shows as mergeable, so merging it merges them. Review its second-pass range first: 31 commits, 15 files, +1725 / −211, of which the checker and its tests are +1607 / −141.

The questions it answers. Where do intent, code, tests and docs disagree, and does each fix reach the whole defect class or only the reported case? What durable content in the frozen records does the graph not carry yet, and where does each piece go? (The roadmap's first entry, the delivery-record cleanup, is blocked on that question.) How far has the product moved from its seed design, and what does the AI direction still need? Does the pull-request record hold anything the graph should?

Three guarantees hold again.

• Delivery facts are derived, never authored. At main, a TypeScript carrier could author hasVerifier, or hide deliveryFacts in a section; now every carrier refuses every delivery-fact name, at every level, in any case.
• A build writes only inside its root, under an honest threat model. At main, sdp build and sdp stamp wrote through a planted link to a file outside the corpus; now every writer rechecks its parent first, and the Spec says these are path checks, not race-free confinement.
• Derivation does not depend on where the engine sits. At main, an engine installed outside this checkout derived this corpus without 105 of its anchors, silently; now it derives the same graph.

Part	Commits	Record
The review	b5e0a8be...3dedfaa4, merged in by 12246171	PR #38, its evidence
The execution	b5e0a8be...c7368f48	PR #39, its evidence
The second pass	4e7e9f95 to bcd2815f, one merge commit per lane	this description, its evidence

The most important changes

1. Writes stay inside the root, across the defect class. The review found that sdp build's staging followed a planted link (L4-01, high); the execution fixed the staging. The second review found the class at more writers, two of them high: sdp stamp wrote through a planted or escaping link (A-c-1), and a watch that refused an escaping generated/ link then deleted the Studio's page set through it (A-c-2). Now every writer, from the build and the publishers' recovery to sdp new spec, projection-suite.mjs and package-artifact generation, rechecks the output parent before each write, rename or delete. A linked generated/ is refused even when it points inside the root. Node has no openat-style calls, so a writer inside the root can still race a check; the rule says the guards detect such a swap rather than prevent it. Stop: the build pipeline's write boundary.

2. No carrier can author a delivery fact, in any case. The honesty check missed hasVerifier (L3-01, high), and the execution gave it one spelling list. The second review found the hole one level down (C-g-1, high): a TypeScript carrier's section records went unchecked, so design.deliveryFacts reached the graph, as did claim, satisfies and the other derived names (S-C-2). The check now reads every key of each section record. As the owner ruled, every carrier now compares delivery-fact names case-insensitively, as Gherkin did (S-C-8), through one matcher: Has-Verifier: in a Markdown rule is refused, while has_verifier stays ordinary text. Stop: machine truth is never authored.

3. The same corpus derives the same graph wherever the engine sits. The extractor trusted only its own builder modules, so an engine installed elsewhere silently dropped 105 anchors (GAP9-1). The owner ruled that when the extraction root is the Protocol package itself, the extractor also trusts that root's src/ids.ts, src/anchors.ts and src/model/code-anchor.ts; lookalikes stay refused. The determinism Spec holds as written; a new test extracts an untouched copy of this repository with an engine outside it and one inside it, and requires equal graphs (GAP9-2). Stop: the anchor model (term "Protocol builder binding").

4. Two reserved Then words, kind and unspecified. A Then slot named kind collided with the generated Outcome's discriminant (L1-03); a Then keyed unspecified could never be bound, silently (D-c-2). The owner ruled to reserve both. Validation reports conformance/reserved-then-slot or conformance/reserved-then-key as an error, and generation withholds the contract with a warning naming the word. An oracle that selects unspecified fails a bound example with its own coverage-gap diagnostic. Stop: executable contracts.

5. "Floor not evaluated" is its own state. The reader evaluates no readiness floor over a Spec whose kind or altitude is unratified (L1-05), and every consumer read that as "below idea" or "nothing blocks" (C-c-5). Such a value comes only from a graph another producer hands to the package. As the owner ruled, the reader marks the Spec floorNotEvaluated, recipes report not evaluated, the Studio shows floor not evaluated with no divergence banner, and the Design Review, census and vault show floor not evaluated (unratified descriptor) (XG-T-1). A kind named constructor no longer crashes a page (F-pr40-claude-1). Stop: the next-rung floor (#notEvaluated).

6. The Node range says what the package needs. The root's require entry loads two ES-module-only dependencies, so it threw ERR_REQUIRE_ESM on Node 20.0–20.18 and 22.0–22.11, which engines: >=20 admitted (GAP5-1). As the owner ruled, engines.node is now ^20.19.0 || >=22.12.0; a package test loads the CommonJS root and fails if it needs an ES module while the range admits a Node that cannot require() one; the docs state the range. Subpath types sit under import, so TypeScript refuses a require that Node refuses (E-c-6). Stop: package.json.

7. A gate for Markdown links. Nothing checked Markdown links, so the cleanup could leave dangling links behind a green gate (S-M38-3). A new test reads every relative link in README.md, AGENTS.md, CONTEXT.md, docs/, jtbd-stories/ and the skills as cmark-gfm reads it, and checks each path and each fragment on a Markdown page against that page's headings (F-sweep2-23). Within those, it skips only the verbatim v0 concept set, so the notes in docs/lineage/ are checked (F-sweep2-10). Stop: test/markdown-links.test.ts.

8. Text scans are linear, and every line ending counts. The execution made one quadratic trim linear (L3-03). The second review found eight more quadratic patterns, from mention punctuation to the Studio's search tokenizer (C-c-1–3, C-g-3, GAP6-1–6-4); each now scans linearly, with a timing test. A lone CR, U+2028 or U+2029 can no longer hide text in any carrier's title, prose, entry or step, or in an @sdp tag line, and TypeScript line numbers count it as the compiler does (GAP6-5, GAP6-6, F-sweep1-4, F-sweep1-11, F-pr40-claude-3). Stop: test/linear-text-scans.test.ts.

9. The test wrapper merges a split run's reports. npm test runs test/cli.test.ts in its own process. pnpm's forwarded -- hid every filter (E-c-3), related and file:line filters went astray (E-c-4, S-E-1), and the CLI pass dropped the caller's options (E-c-5); the wrapper now routes each form as Vitest reads it. For report files the owner ruled to merge (F-sweep2-20): each pass writes a blob report, and vitest --merge-reports writes the one report the caller asked for. Coverage stays refused: no coverage provider is installed. Stop: vitest-test.mjs.

10. The accepted inventory, and how the cleanup lands it. The cleanup entry, spec:protocol.delivery-record-cleanup, is held at idea by #durableContent: what in plans/, reviews/ and the concept documents does nothing carry yet? The inventory answers it: most is carried or history; fourteen pieces need a home, four live dependencies must move, and 21 Markdown links and the handbook's and glossary's pointers must be repointed. The owner accepted it with row rulings. It lives in removable evidence, so the cleanup arc lands it in its first commit, which closes the question, as its provisional now says (J-c-3). Stop: the inventory's rulings.

Shown to a person. The second pass moves no human view of this corpus: the floor label needs an unratified descriptor, which no carrier can state. The execution moved the Studio's intent panel: each gathered intent names its action (Q6a), and a closed Preview memo shows the memo before it is copied (Q6b). Compare the panel's page in the Studio preview of main (PR #36's preview) with this pull request's. Each preview's stamp names a merge commit whose tree is exactly main's and this head's. These are live previews that track each push, not pinned images.

How the work ran

• The review. Twelve lanes, each with a reader from one model family (Claude or GPT) and a verifier from the other who tried to refute each finding: 73 items, 55 confirmed (2 high, 26 medium, 27 low), 16 refuted, 2 unsettled, both since settled by the owner. The dossier gives each lane's scope and limits.
• The execution. Nine lanes in two passes (C, W, H1, H2, P, D, then S, L, K), each implemented by one family and reviewed by the other until approved.
• The second review. Fifteen lanes over PRs docs(articles): publish the product and projection narrative #34, review: the whole repository, and the durable-content inventory #38 and fix: execute the whole-repository review's findings #39, twelve read independently by both families and three by one: 120 findings, merged into 76 distinct defects, each verified by the family that did not find it: 70 confirmed, 6 refuted. Two sweeps re-read every report for anything said but not recorded (35, then 9, more confirmed).
• Four fix rounds. Round 1 ran twelve lanes: X1–X9 here, Y1 and Y2 on PR docs(articles): publish the product and projection narrative #34's branch, and Z on PR review: the whole repository, and the durable-content inventory #38's branch, now carried here; round 2 took the second sweep's follow-ups. An audit checked all 114 confirmed ids against the tree. A verification of the fixed tree confirmed 42 of 54 new items for round 3; a closing review of round 3 confirmed all 16 of its items for round 4, which ran three lanes here and one on PR docs(articles): publish the product and projection narrative #34's branch. Each lane was implemented by one family and reviewed by the other until approval; reviewers re-measured each proof by reverting its fix.
• In all, the second review recorded 199 findings: 27 refuted and 172 confirmed (3 high, 23 medium, 146 low), 133 carried here and 39 on PR docs(articles): publish the product and projection narrative #34's branch. Severities are the verifier's. Two of the three high findings existed at main (A-c-1, C-g-1); the third came with the execution's own refusal path (A-c-2).

Working record

• Branch fix/second-review-pass; base main b5e0a8be. It opened on PR fix: execute the whole-repository review's findings #39's branch at c7368f48, and a merge commit brings in PR review: the whole repository, and the durable-content inventory #38's branch.
• The owner's working rules. Review the whole repository, then execute the confirmed findings and rulings, with Claude and GPT as two model families: each finding is verified by the family that did not find it, nothing a reviewer reports may be lost, and each fix lane is implemented by one family and reviewed by the other. Then review and fix PRs docs(articles): publish the product and projection narrative #34, review: the whole repository, and the durable-content inventory #38 and fix: execute the whole-repository review's findings #39 again, leaving no stone unturned. The owner reviews everything here. Commits, push and this draft are authorized; the owner merges.
• No arc Pack. The arcs execute findings and rulings and add no corpus intent of their own; their corpus changes are measured with the design diff (recipe 35) from main.
• Evidence. evidence/pr38/, evidence/pr39/ and evidence/pr40/ are arc evidence, removed before merge; the links are pinned to commits the merge keeps reachable. The full per-id record, with each finding's verdict, where it was handled and the commits that name it: evidence/pr40/findings.md.

What changed, by area

Outcomes: fixed (with a test that fails without the fix), applied (as ruled or written), adapted (as a recorded choice or later ruling reshaped it), description, recorded. Proof names files under test/; the ledger names each test. "#39" marks the execution.

Area	Findings	Outcome	Proof	What changed
Writes	L4-01 (high, #39), A-c-1 (high), A-c-2 (high), A-c-3, A-c-4, A-g-1, A-g-5, L4-02 (#39)	fixed (A-g-1 a mitigation)	cli-write-boundaries	exclusive staging; every publication, recovery, stamp, script and package-artifact mutation rechecks its parent; a linked generated/ is refused; a case collision refuses publication
Writes	A-g-6, F-sweep1-6, F-sweep1-17, F-sweep1-18	fixed (F-sweep1-17 adapted)	cli-new-spec, cli	scaffold and registrar parents created through held directories; each recheck named; --help says where PATH must resolve
Writes	GAP1-1, S-A-5, F-pr40-claude-2, F-sweep1-1, GAP8-1	fixed (F-pr40-claude-2 adapted)	cli-write-boundaries, studio-s4	vault placement checked before the build, each sync and the last cleanup; a default vault resolves inside the root; a failed stamp leaves no staging file; the preview discards the old page set
Writes	A-c-5, A-c-6, claude-side-T-5, F-sweep1-22	applied	oracle pins, cli-write-boundaries	the rules that carry the boundary; the build's hooks documented
Discovery and watch	Q1 (ruling, #39), Q1n (#39), B-c-1, claude-side-T-4, L4-04 (#39)	applied, fixed (L4-04 adapted)	exclude-diagnostics, validate-watch	no link followed below the root; a link warns only when its entry could hide a carrier or anchor; only a watch resolves its root's physical path
Discovery and watch	B-c-2, B-g-2, F-sweep1-3, F-sweep1-16	fixed	validate-watch	a watch reruns when a warned link comes or goes; a replaced directory, the root included, is dropped or rewatched
Honesty	L3-01 (high, #39), C-g-1 (high), S-C-2, S-C-8 (ruling), F-sweep1-25, VR2-T-2	fixed (S-C-2 adapted: refused in validation), applied	validators, extract, markdown-reifier, gherkin-reifier	one case-insensitive matcher; every key of a section record checked; @deliveryFacts refused; refusal wording
Carriers	L1-04 (#39), C-c-4, C-g-2, S-C-1, F-sweep1-7, F-sweep1-8, F-sweep1-9	fixed, applied	markdown-reifier, oracle pins	prose and list text refuse headings, fences, short setext underlines and lone list markers; the key and reserved-property rules say what is read, and where
Line endings	GAP6-5, GAP6-6, F-sweep1-4, F-sweep1-11	fixed	gherkin-reifier, anchor-forms, extract	a lone CR, U+2028 or U+2029 is a line ending everywhere it is read
Linear scans	L3-03 (#39), C-c-1, C-c-2, C-c-3, C-g-3, GAP6-1, GAP6-2, GAP6-3, GAP6-4, F-sweep1-23, F-sweep1-24	fixed (GAP6-2, GAP6-3 adapted)	markdown-ascii-trim, linear-text-scans, codegen	linear scans; example spaces decode over one index
Runtime and contracts	L1-01, L1-02, L1-03 (ruling), L5-01, L5-02 (#39)	fixed	codegen, self-hosting-graph	one handler per shared key; an unspecified oracle matches no Then; kind reserved; two suites bound as verifiers
Runtime and contracts	D-c-2 (ruling), D-c-3, D-c-4, D-g-2, GAP7-1, GAP7-2, F-c-1, F-sweep1-2	fixed	codegen, runner, testing, studio-s5	unspecified reserved; runtime imports aliased; no inherited name stands in for a handler, Condition, parameter or label
Floor not evaluated	L1-05 (#39), C-c-5, XG-T-1 (rulings), F-pr40-claude-1, F-sweep2-12	fixed	reader, recipes, studio-floor-not-evaluated, projections-floor-not-evaluated	floorNotEvaluated, labelled everywhere a floor is shown; own-key lookups and tallies
Builder trust	GAP9-1 (ruling), GAP9-2	fixed	anchor-trust, cli, package-smoke	the Protocol root's own builders trusted
CLI and package	L2-3, L2-5, L9-F2, C1 (#39), E-c-1, E-c-2, E-g-3, E-c-6, GAP5-1 (ruling), M38-c-7	fixed, applied	cli-q, studio-s3, package-smoke, bootstrap, check-temporal	parse errors placed where the engine reads them; typed CommonJS root; the Node range; the guard sweeps DECISIONS.md
Test wrapper	L4-03 (#39), E-c-3, E-c-4, E-c-5, S-E-1, claude-side-T-2, claude-side-T-3, F-sweep2-20 (ruling)	fixed (E-c-5 adapted, then superseded by the ruling)	self-hosting-duplicate-ids, with real Vitest runs	routing as Vitest reads it; options reach both passes; merged reports
Test infrastructure	S-I2-3, S-N-2, F-sweep1-13, S-J-1, S-N-1, F-sweep1-14	fixed	package-smoke, cli, integration-review, vault-probe; budgets have no deterministic proof	tests independent of the npm cache, ancestor configuration and temporary directory; the snapshot race; budgets
Docs and vocabulary	L6-1, L6-4, L6-6, L2-6, L9-F1, L9-F4, L9-F5 (#39), H-c-1, H-c-3, H-c-4, S-H-1 (ruling), G-c-2 (ruling), G-c-3, J-c-3	fixed, applied	self-hosting-graph, recipes, skills, checkout-v1; job stories have no test	docs agree with the graph; carrier at both levels; #durableContent's provisional
Link gate	S-M38-3, F-sweep2-23, F-sweep2-10	fixed	markdown-links	paths and fragments; the lineage notes checked
Corpus rulings (#39)	P1–P7 (L2-1, L2-2, L6-3, L12-01), Q1–Q6, K1–K3 (L3-02, L9-F3, L10-9, L11-01, L12-02), Q6a, Q6b, SP1	applied, adapted, fixed	oracle pins, recipe 20, studio-s5, studio-slice	the accepted proposal; three idea captures; JS-F1's job; labelled intents; #memoPreview; #exportOmitsSlice
Review evidence (PR #38)	M38-c-2, M38-c-3, M38-c-6, M38-c-8, M38-c-9, GAP13-1	adapted, fixed	check-temporal; the rest is evidence prose	one closure condition; row R12; the 21 links listed; 213 pinned links counted; exemptions pinned to the root
Round 4: runtime, recipes	F-sweep3-2, F-sweep3-1, F-pr40-claude-4, F-sweep3-3	fixed (F-sweep3-2 adapted), applied	runner, recipes, oracle pins	dispatch boundaries hold in any realm, stated as #handlerDispatch; recipes 5, 21 and 31 tally foreign readiness exactly
Round 4: carriers	F-pr40-claude-3, F-sweep3-5, F-sweep3-6	adapted, fixed, applied	markdown-reifier, import-round-trip, gherkin-view, gherkin-reifier	one line-ending law in every carrier; Gherkin refusals name their Spec; rule wording
Round 4: wrapper, watch	F-sweep3-8, F-sweep3-9, F-sweep3-14	adapted, applied, adapted	self-hosting-duplicate-ids, validate-watch	@vitest/ui checked first; configuration read as Vitest reads it; fewer watcher rebuilds
Descriptions	D-c-1, J-c-2, J-c-5, J-c-6, N-c-5, F-c-3, S-C-4, S-I1-2, F-audit-4, F-audit-8, F-audit-9, F-audit-14, F-sweep2-18	description	—	PR #39's description, carried forward corrected here
Descriptions	M38-c-4, M38-c-5, M38-c-10, S-merge-5, claude-side-T-1, F-sweep2-24	description	—	PR #38's description, carried forward corrected here
Records	N-c-1, F-sweep1-27, F-sweep2-17, F-sweep3-7, F-sweep3-10	recorded	—	the sections below; for F-sweep3-10, bcd2815f adds the later rounds and the ledger to the evidence

Where the review's other findings went. The concept-document drift (L6-2, D1, L6-5) and the record findings (R1–R12, L8-01, L8-02) go to the delivery-record cleanup. The eight article corrections (L10-4, L10-5, L10-7, L10-8, L10-11, L11-02, L9-F6, L9-F7) are in PR #34, with the 39 second-review findings against that pull request.

New in the corpus

The design diff (recipe 35) from the main graph, built by main's own CLI, measured at the head: 3 Specs added, 43 changed, none removed; 1 Pack changed; 9 bindings changed, each gaining a verifier. What kinds of change landed:

• Three idea captures, each with a blocking deferral, in pack:self-hosting-v1: AI-assisted reading, bound-file freshness and combination coverage.
• Answered questions become rules. Cited-line text records #lineRecorded and design entry lineBound; the slice page gains #exportOmitsSlice; the intent panel gains #memoPreview. The cleanup gains #inventoryFirst and keeps #durableContent blocking, its provisional naming the acceptance and the closing commit.
• New rules on ready Specs: #entryAddressResolution (reader), #reservedThenSlot and #reservedThenKey (slot notation, executable contracts, runnable modules), #notEvaluated and #floorNotEvaluated (the floor and the answer shapes), #writeBoundary and #exclusiveStaging (build pipeline), #defaultInsideRoot (vault), #scaffoldWriteBoundary (authoring on-ramp), #physicalLines (Gherkin authoring), #lineEndings (Markdown authoring), #handlerDispatch (example runner).
• Publication rules on the frozen projections' Specs: #portablePublication and #safeCleanup on the Design Review, Mermaid, Gherkin and wholesale-rewrite Specs, and a qualified posture on the projections model (delegated, below). The census, the readiness banner and the vault notes gain the ruled floor label.
• Rules reworded to match the code: discovery's link warning, Markdown block shapes and keys, reserved names in every carrier, the trusted builders, the stamp, the watch, and every Spec that reports a floor.
• Bindings: test/cli-write-boundaries.test.ts verifies six Specs, and three regression suites became verifiers (the Markdown body grammar's, the structural anchors' and the scaffolder's).
• Round 4 added #lineEndings, the one line-ending law for every carrier, with matching rules on the envelope contract and the Gherkin view; #handlerDispatch, which executable contracts' planDispatch now points at; and clearer #blockShapes and #keyedEntryShape.

The decision registry gains no row; the carrier ruling (MD-18), the Pack syntax ruling (MD-25), the Gherkin carrier option (MD-27) and planning truths live in ruled graph homes (MD-33) now name the decisions that narrow them.

Rulings

The owner's rulings

Ruling	One line	Carried by
Accept the seven-change proposal	1–3 repair agent-surface drift; 4–7 record answers accepted in PR #36	reader, cited-line text
Discovery never follows a link (Q1)	a link found warns; a root named explicitly resolves first	excludes
The panel hands off scope, not a budget (Q2)	the job changes, not a Spec	JS-F1
Three captures (Q3, Q4, Q5)	AI-written presentation, bound-file freshness and slot combinations	AI-assisted reading
The hand-off preview splits (Q6)	the missing action label is a defect; a collapsed memo preview, with a rule and a UI entry	intent panel
The Studio build composes joins (U2)	refused is a page that recomputes a join, or a second graph	Spec Studio #noSecondGraph
PR #32's delegated rulings stand (U1)	as the graph carries them; a later narrowing wins	no edit
PR #38's delegated choices	seven ratified, the guard's evidence/ exemption among them; one superseded	check-temporal.mjs
The inventory answers #durableContent	with row rulings: R1, R5, R7, R9, R11, L8-01, L8-02 take their proposed homes; R4 becomes spec:carrier.typescript-authoring; R2 goes to the cleanup's pull-request record, never a Spec; DECISIONS.md stays; the plan-reading gate checks retire with the folders; measure.mjs moves beside its test	the rulings, cleanup
The cleanup comes first; PR #36's provisional answers stand	recorded by proposal changes 4–7	cleanup
A Pack export leaves out the slice page	the question becomes a rule	slice page
The eight article corrections	accepted	PR #34
Reserve kind in Then slots (L1-03)	an error and a withheld contract; the Outcome keeps its flat shape	slot notation
Reserve unspecified as a Then key (D-c-2)	mirrors kind, with its own coverage-gap diagnostic	runnable modules
Narrow JS-F1 criterion 7 (G-c-2)	"When the panel offers a memo, it previews the exact memo before it leaves."	JS-F1
PR #34 stays pinned to b5e0a8be (L34-c-9, L34-g-3)	the first refresh pull request after this one re-pins and regenerates the evidence	PR #34's description
A floor not evaluated is its own state (C-c-5)	in the reader, recipes 7, 9, 11 and 55, and the Studio	next-rung floor
…in the frozen projections too (XG-T-1)	a corrective label in the Design Review, census and vault; no new decision Spec	census page
Trust the extracted Protocol root's own builders (GAP9-1)	lookalikes refused; closes GAP9-2 and PR #38's question on dropped anchors	anchors
Narrow the Node range (GAP5-1)	^20.19.0 || >=22.12.0, a CommonJS load gate, and the docs	package.json
Gloss carrier at both levels (S-H-1, H-c-4)	the format; "a carrier file" is one file in it	CONTEXT.md
Delivery-fact names in any case (S-C-8)	every carrier, as Gherkin did; underscore forms stay ordinary	authored honesty
Merge a split run's reports (F-sweep2-20)	coverage refused while no provider is installed	vitest-test.mjs
One pull request for three arcs	PRs #38 and #39 close in favour of this one; PR #34 stays separate	—

Delegated choices awaiting the owner's review

Choice	One line	Where
Rules on the frozen projections' Specs (A-c-5, A-c-6)	#portablePublication and #safeCleanup on Specs the shipped projections stay frozen (MD-32) protects, with no refining decision; pages are unchanged, but a case collision exits 1	Design Review
A linked generated/ is refused even inside the root (A-g-1)	and the threat model is path checks, not race-free confinement	build pipeline
A default vault must resolve inside the root (S-A-5, F-pr40-claude-2)	one linked outside or nowhere is refused; a named vault may sit outside	vault
The stamp names its staging file (F-sweep1-1)	#stampVerb said the stamp touches no other file	sharing
A replacing directory is watched in its place (F-sweep1-3, F-sweep3-14)	the root included; a root that no longer resolves ends the watch with exit 1; a directory is rebuilt only when replaced or reporting itself	validate watch
Gherkin physical lines (F-sweep1-11)	titles, steps and description lines refuse a lone CR, U+2028 or U+2029	Gherkin authoring
Reserved properties and nested content (F-sweep1-9)	clarifies existing behavior on a ready Spec: a Model term or a key inside a Design value stays content	authored honesty
Keys and lone list markers (F-sweep1-7, F-sweep1-8)	a key is a letter then letters or digits, and a hyphenated delivery-fact name reads as one; a lone list marker is refused	Markdown body grammar
The scaffolder's boundary (claude-side-T-5, F-sweep1-17)	each recheck names its checkpoint; a test-only hook	authoring on-ramp
The floor label's name and reach (C-c-5)	floorNotEvaluated lists the descriptors; every recipe reporting a floor says not evaluated, beyond the four ruled	answer shapes
Inherited class methods are step handlers (F-sweep1-2, F-sweep3-2, F-sweep3-1)	a class instance's methods dispatch, and so do a built-in parent's (Map, Array, Promise); any realm's Object.prototype and Function.prototype, and an inherited constructor, are refused	example runner #handlerDispatch
One line-ending law across the carriers (F-pr40-claude-3)	Markdown also refuses U+2028 and U+2029, at the physical line; TypeScript refuses a Spec or Pack whose kept strings hold one; the Gherkin view, a frozen projection, folds every line terminator to a space	Markdown authoring #lineEndings
The wrapper's routing and reports (L4-03, E-c-4, F-sweep2-20)	list, init and bench run in one spawn; related always splits; a file:line filter is read as its file; a blob report is refused, since Vitest cannot merge into one; an HTML report stops before any pass while @vitest/ui is missing	vitest-test.mjs
The link gate's scope (S-M38-3, F-sweep2-10, F-sweep2-23)	not carriers, plans/, reviews/, examples/, explorations/, evidence/ or the v0 concept set; fragments match leniently; line anchors stay unchecked	markdown-links.test.ts
A lineage note's dead links repointed (F-sweep2-10)	the dated annotation scout's links to two removed decision Specs point at 35539a0d, the last commit holding both	docs/lineage/
The preview workflow discards the old page set (GAP8-1)	so only a page set this run published is stamped	studio-preview.yml
Slow-test budgets (S-N-1, F-sweep1-14)	raised to hold under load, with no deterministic proof	test/
Proposal change 5 stays as accepted (PR #38)	the provisional it produced names the acceptance (J-c-3)	proposal.json
L4-04 narrowed to the watch root; L4-02's collision refused, not renamed (#39)	output keeps the named path and page names	src/cli/
Q1's warning narrowed (lane L, B-c-1)	departs from "each link it finds gets one warning": a non-candidate file link, and a directory link named like a skipped directory, stay silent	excludes #canonicalDiscovery
Q6a's labels (#39)	one label map; an unknown action shows its raw string	intent-panel.ts
StepFn gains a required third parameter (L1-01)	public on ./runner: code that calls, wraps or retypes a StepFn must pass the step	runner/index.ts
The root's require condition gains types (L9-F2); the delivery-fact spellings stay internal (L3-01)	the published export map changes; the exports do not	package.json
Reserved-word findings are errors; generation withholds (L1-03)	the ruling named no severity; every generation finding is a warning; D-c-2's ruling then mirrored this	executable contracts
C1 fixes the exemption only; the carrier ruling (MD-18) names its narrowers too (L6-4) (#39)	the temporal guard's law gets no Spec	check-temporal.mjs, DECISIONS.md
Q3's relation types (#39)	refines the projections model; dependsOn three Specs	AI-assisted reading

Later rulings supersede two of the execution's delegated rows: C-c-5 replaces "an invalid altitude skips floor evaluation" (L1-05), and G-c-2 replaces "JS-F1 criterion 7 capitalized".

An honest record

• Review-process text in pushed commit messages. 52262422 here, and efc5f7d9 and 8fc1802b on PR docs(articles): publish the product and projection narrative #34, carry working notes that point at nothing durable (F-sweep2-17, F-sweep2-9). They stay, since rewording would rewrite pushed history; their change descriptions and finding ids are the record.
• A lint-red stretch. 29810f11 and 78e9e706 leave npm run lint red until f3bf5f64 repairs it (F-sweep1-27); a bisect that lints should skip them. On PR docs(articles): publish the product and projection narrative #34, 43d63901, 4c614845 and 5efd7312 leave format:check red until 0389c560 (F-sweep3-13).
• The opening commit's CI run failed. The run on 4e7e9f95 failed test/cli.test.ts "views the self-hosting corpus from the default repository root" at 15.1 s against its 15 s budget. a9f94171 raises it to 120 s, with six other budgets of the same class (S-N-1).
• A flaky gate at PR fix: execute the whole-repository review's findings #39's head. The vault watch example's snapshot helper could race the sync's staging rename and fail once with ENOENT on a .sdp-tmp file; a51bf5c6 fixes it (S-J-1).
• The base moved. The branch opened on PR fix: execute the whole-repository review's findings #39's branch and now targets main; the opening commit's message still names the earlier base.

Out of scope, and left open

• For the cleanup arc: the record findings and the concept-document drift; R6's re-entry conditions, or dropping R6; R8's home; R12, keep it with a home or drop it; the new location of the plan-46 prototypes, which two Specs cite. None holds #durableContent open; each is settled before the record that holds it is removed. Nothing in plans/ or reviews/ is deleted here.
• Open by design: register-input's three questions (registerEntry and inputForm, blocking, and textLinks), for its design pass at roadmap place 3; designColumnsSufficiency on the design-maturity Spec, which waits for the first adopter's design passes.
• The temporal guard over docs/lineage/ (F-sweep3-7) needs a decision. The guard exempts the whole folder as "verbatim historical exhibits", but authored notes live there: the folder README, and PR docs(articles): publish the product and projection narrative #34's annotation context and dated scout. Narrowing it fails both pull requests on their dated lines, so the exemption carries weight while its stated reason is false. Either keep it and say it covers dated historical notes, or narrow it to v0-design/, exempt the scout by name and reword the two dated lines; either way a test pins the boundary.
• An unratified stated readiness, such as done in a foreign graph, is outside both floor rulings and is left as it was.

Merging beside PR #34

The owner chose the order: this pull request merges first, and PR #34 then merges main into its own branch and merges once CI passes there. Before this merge, the combined tree (this head 93cf4194 with PR #34's head 22c49262 merged on top) passed npm run check (135 files and 4,137 tests, plus 102 in the CLI pass) and npm run check:articles (11 blocks at b5e0a8be). That was a local run; the gate of record is PR #34's CI on its updated branch.

Git reports one content conflict, at the top of README.md, and it is the same in either order. This branch moves the README's Carrier rule: paragraph under a new ## Carriers section and puts a newcomer's sentence in its place; PR #34 keeps the paragraph inside its new product-story section. PR #34's branch resolves it in this order:

1. this branch's sentence ("You write Specs as Markdown files…") directly after the opening paragraph;
2. PR docs(articles): publish the product and projection narrative #34's ## Read the product story, without its Carrier rule: paragraph;
3. the existing "Read the [concept]…" line.

Check: grep -c 'default to Markdown' README.md prints 1 (N-c-1), and prettier --check README.md passes. Once both have merged, the link gate reads PR #34's articles and lineage notes too, and the scout's repointed links follow PR #34's rename of that file. The first refresh pull request after this one re-pins the articles, as the owner ruled; it needs this pull request's section-level deliveryFacts refusal (S-L34-2).

Try it

git fetch origin fix/second-review-pass && git checkout fix/second-review-pass
npm ci && npm run build

# 0 errors and the three pinned honesty/gaps warnings; note the "specs · packs · anchors" line
pnpm --silent sdp validate . --exclude explorations --exclude examples --exclude test/fixtures/import/parity

# The open-question register: the three captures, and #durableContent's provisional
pnpm --silent sdp:q "$(cat dist/recipes/20-open-question-register.js)"

# A linked generated/ is refused even inside the root: "unsafe output parent ... is a symbolic link"
mkdir -p /tmp/sdp-try/corpus/inside && ln -s inside /tmp/sdp-try/corpus/generated
cp specs/consumers/ai-assisted-reading.sdp.md /tmp/sdp-try/corpus/ && node dist/cli/sdp.js build /tmp/sdp-try/corpus

# A delivery fact in any case: extract/reserved-property for "Has-Verifier"
mkdir -p /tmp/sdp-try/fact && printf '%s\n' '---' 'id: spec:try.fact' 'kind: behavior' 'altitude: story' 'readiness: idea' 'relations: {}' '---' '# A hand-authored fact' '## Intent' '- outcome: Show the refusal.' '## Rule' '- Has-Verifier: the suite' > /tmp/sdp-try/fact/f.sdp.md
node dist/cli/sdp.js validate /tmp/sdp-try/fact

# A Then keyed unspecified: a conformance/reserved-then-key error; the space contract is withheld
mkdir -p /tmp/sdp-try/unspec && printf '%s\n' '---' 'id: spec:try.unspecified' 'kind: behavior' 'altitude: story' 'readiness: idea' 'relations: {}' '---' '# A reserved Then key' '## Intent' '- outcome: Show the refusal.' '## Example space' '```gwt-vocabulary' 'Given an order' 'When the order is placed' 'Then unspecified' '```' > /tmp/sdp-try/unspec/u.sdp.md
node dist/cli/sdp.js validate /tmp/sdp-try/unspec

# A foreign graph with an unratified altitude: [ 'altitude' ], then "| floor not evaluated (unratified descriptor) | 1 |"
node --input-type=module -e '
import { createReader, extract, renderCensus } from "./dist/index.js";
const { graph } = extract({ root: "examples/checkout-v1" });
const id = "spec:orders.order-placement-flow";
const foreign = { ...graph, nodes: graph.nodes.map((n) => (n.id === id ? { ...n, altitude: "planet" } : n)) };
console.log(createReader(foreign).specContext(id).floorNotEvaluated);
console.log(renderCensus(createReader(foreign)).find((p) => p.path === "index.md").content.match(/^\| floor not evaluated.*$/mu)[0]);
'

# The same graph wherever the engine sits: the copy prints the same "specs · packs · anchors" line
mkdir -p /tmp/sdp-try/copy && git archive HEAD | tar -x -C /tmp/sdp-try/copy
node dist/cli/sdp.js build /tmp/sdp-try/copy --exclude explorations --exclude examples --exclude test/fixtures/import/parity

# The link gate; one merged JUnit report from the split run (the second command runs the whole suite)
npm test -- test/markdown-links.test.ts
npm test -- --reporter=junit --outputFile=/tmp/sdp-try/report.xml

# The design diff from main, against a graph built by main's own CLI
git worktree add ../sdp-main b5e0a8be && (cd ../sdp-main && npm ci && node dist/cli/sdp.js build . --exclude explorations --exclude examples --exclude test/fixtures/import/parity && cp generated/graph.json ../main-graph.json)
pnpm --silent sdp:q "$(cat dist/recipes/35-the-design-diff.js)" --params base=@../main-graph.json

The quoted outputs were re-measured at bcd2815f, whose engine matches f5f7fd5b's build (bcd2815f changes only evidence/). The link gate ran inside npm run check. The JUnit command, re-run at bcd2815f, wrote one merged report of 4,112 tests in 135 files with no failure, the CLI pass included. The intent panel is in the live previews above: under rule.memoPreview, "Propose new text", "Gather", then open "Preview memo" in "Composing (1)".

Feedback wanted

• Rules on the frozen projections' Specs. #portablePublication and #safeCleanup change no page but add rules to four Specs that the shipped projections stay frozen (MD-32) protects. Do they belong there, or in a decision that refines that ruling?
• The write boundary's threat model. Path checks before each mutation, a refused linked generated/, and an identity check that detects a swap rather than prevents it. Is that the right line for a CLI with no openat-style calls?
• Where a link warns. Only when the entry it stands for could hide a carrier or anchor. Is that Q1's intent, or should every link warn and this repository exclude its own?
• The default vault. Refused only when it resolves outside the root, while a named vault may sit outside. Is that the placement the vault Spec means?
• The temporal guard over docs/lineage/: which boundary?

Measured at the head

Re-run the recipes rather than inheriting these numbers. The base column was re-measured for this description with main's own CLI over an archive of b5e0a8be; both columns use the corpus exclusions.

Measure	Base b5e0a8be	Head
Specs · Packs · anchors	456 · 9 · 607	459 · 9 · 611
Graph nodes · edges	1072 · 2978	1079 · 2996
sdp validate	0 errors · 3 warnings (the pinned honesty/gaps)	0 errors · the same 3 warnings
Build backlog (recipe 1)	2: spec:consumers.edit-model, spec:consumers.intent-composition	the same 2
Drift alarm (recipe 2)	0	0
Open questions (recipe 20): total · blocking · non-blocking	16 · 10 · 6	15 · 12 · 3
Deferrals · holds · Specs with questions · malformed	6 · 4 · 12 · 0	9 · 3 · 12 · 0
Mention audit (recipe 22): unbacked · unresolved	0 · 0	0 · 0
npm run check	not re-measured	at f5f7fd5b (the head less its evidence commit): every step through check:example passed, 134 files and 4,010 tests plus 102 in the CLI pass; the last step, preflight, flagged the then-uncommitted evidence files bcd2815f commits
npm run check:git-install	not re-measured	pass at f5f7fd5b: npm and pnpm clones, npm archive
CI	—	pass on bcd2815f: push and pull-request runs
Diff from main	—	137 commits (26 merges); 217 files, +17,042 / −1,040; without evidence/, 201 files, +12,884 / −1,040

Next

The next work, from the roadmap (recipe 55, rendered at the head)

Roadmap: The Protocol's roadmap pack:roadmap

To state means the floor allows ready and a person still judges the statement. Built means every Spec of the entry's work states ready and has an implementation binding; it reports no passing test and no deployment.

The Protocol's next work, first entry first. The order is a preference: it dates, gates and sequences nothing, and what each entry still needs is read from the graph each time.

First the cleanup that removes the frozen plans, review reports and superseded concept material, as the roadmap and the pull request record make possible. Then the two steps a reader of the Studio's pages could not take, the words of a test line that checks a declaration and an outside register's entries, and then the composed-intent model the build backlog holds. What the first adopter's mass update asks for joins when the owner selects it; a capture landing adds nothing here. An entry leaves by an edit to this manifest once its selected work is complete or no longer belongs among the next work.

To design

Place	Entry	Next	Purpose and reach
1	The Protocol removes its frozen plans, review reports and superseded concept material spec:protocol.delivery-record-cleanup	itself idea → scoped, nothing holds it; blocked by spec:protocol.delivery-record-cleanup#question.durableContent	Remove plans/, reviews/ and the superseded concept material from the tree, after moving into the graph the durable content they still hold, so the repository keeps one delivery record: the graph, git and the pull requests.
3	An outside register is a named Studio input, so a cited token links to its entry spec:consumers.spec-studio.register-input	itself scoped → defined, held by no-blocking-open-questions; blocked by spec:consumers.spec-studio.register-input#question.registerEntry, spec:consumers.spec-studio.register-input#question.inputForm	A cited token reads as bare text on every page, and the register that defines it is neither linked nor shown.

To state

Place	Entry	Earlier waiting statements	Purpose and reach
2	A pinned declaration shows the words of each test line that cites it spec:consumers.spec-studio.cited-line-text	No preceding waiting statement in the statement queue.	The declarations page names each citing line by file and line only, so the reader leaves the page to read it, and has no route at all when the page set carries no source link.

To build

Place	Entry	Purpose and reach	Rests on
4	Views compose scoped intent instead of patching canonical source spec:consumers.edit-model	Let a view frame a requested change without giving derived surfaces a direct write path to canonical source.	nothing unbuilt

Waiting on the owner

• spec:protocol.delivery-record-cleanup#question.durableContent (blocking), place 1: Which content in plans/, reviews/ and the concept documents does no Spec, test, decision or recipe carry yet, and where does each piece belong?
• State spec:consumers.spec-studio.cited-line-text: it states scoped and its floor allows ready; place 2
• State spec:protocol.delivery-record-cleanup: it states idea and its floor allows scoped; place 1

Deferred, with their triggers

• spec:carrier.sdp-import.feature-adapter#question.featureAdapter: Which bare .feature constructs can an import adapter preserve honestly? Trigger: an adopter migrates a gen-1 .feature corpus into the Protocol
• spec:consumers.agent-surface.combination-coverage#question.combinationCoverage: Which combinations of slot values should coverage enumerate and report? Trigger: a design pass asks which combinations no example exercises
• spec:consumers.ai-assisted-reading#question.aiPresentationBoundary: If an assistant generates explanations or arrangements beyond selecting deterministic readings, what contract keeps its output distinct from graph claims and makes the exact reading reviewable, including how the exact output is retained with its model, prompt and inputs? Trigger: the UX design pass measures that assistant selection or a generated explanation helps a named review task
• spec:consumers.impact-graph.symbol-reach#question.symbolIdentity: Which language-neutral identity and extraction boundary can support exhaustive symbol reach without freezing a single compiler's representation into the Protocol? Trigger: a reader needs symbol-level reach that file-level imports and curated anchors cannot give
• spec:consumers.projections-model.roadmap-views#question.releaseReader: Does a reader need release or baseline views that the roadmap, the recipes and git tags do not serve? Trigger: the first published release
• spec:consumers.reference-projection#question.namedOutsideReader: Does a human reader outside the repository need a reference projection the four shipped roots do not serve? The taxonomy row is not that reader. Trigger: a named consumer outside the repository, with a way to measure that demand inside the arc that builds the projection
• spec:consumers.structural-mermaid#question.namedStructuralReader: Does a reader need diagrams rooted at code units, which census tables and Spec-rooted one-hop Mermaid do not serve? Trigger: a named reader with that need, plus a decision that refines spec:decisions.shipped-projections-frozen for the affected Mermaid rendering or supersedes that ruling whole
• spec:observation.run-evidence.bound-file-freshness#question.freshness: When should changes to bound files make recorded run evidence stale? Trigger: a reader cannot tell whether a run at another commit still applies, because nothing says whether its bound files changed since
• spec:observation.runtime-overlay#question.observationIdentity: Which external observation identity and freshness boundary is small enough for the graph while still supporting an honest observed delivery fact? Trigger: An adopter needs runtime measurements linked to measurable Spec targets and can supply an observation identity and freshness policy for review.

Why this order (the owner's reasons, unchanged since PR #36; the memo does not carry them):

Place	Entry	The owner's reason
1	spec:protocol.delivery-record-cleanup	A dedicated pull request removes plans/ and reviews/ once PR #36's precedent has landed
2	spec:consumers.spec-studio.cited-line-text	Closes a step a reader could not take in PR #35's reader tasks
3	spec:consumers.spec-studio.register-input	Closes the other such step
4	spec:consumers.edit-model	The composed-intent model is the whole build backlog

Re-entry: the cleanup at place 1 opens once this pull request merges. Its first commit lands the accepted inventory, as ruled, and closes #durableContent; R6, R8, R12 and the prototypes' new location are settled within that arc, each before the record that holds it is removed.

Refused, durably (PR #36's list, with the Studio item as U2 worded it): a rung stated by tooling; priority as a date, a schedule, a gate or a new relation; a roadmap that lists every feature; a slice or a capability as a new grouping primitive; an "over-designed" check; a diagram derived from prose; an anchor added for coverage; a tool that edits a carrier itself; a count, date or plan number inside a Spec; a Studio page that recomputes a join, or a second graph.

The owner's decisions this pull request asks for:

1. Review the delegated choices, above all the rules on the frozen projections' Specs, the write boundary's threat model, the narrowed link warning, StepFn's required parameter and the export-map change.
2. Choose the temporal guard's boundary over docs/lineage/ (F-sweep3-7).
3. Rule on R12, keep or drop, before the cleanup removes concept 01.
4. Review PR docs(articles): publish the product and projection narrative #34's second-pass changes on its own branch and merge it first; then merge this pull request with the README resolution above.

This pull request completes no roadmap entry, so the manifest is not edited.

🤖 Generated with Claude Code

Confidence Score: 5/5

The reviewed changes appear safe to merge; no actionable defect was established.

Summary

This PR hardens writes, carrier checks and generated contracts, then updates the Specs, docs and tests to match.

• Build and publishing commands check output paths before changing files.
• Carriers cannot author delivery facts, whatever case they use.
• Readers show when a readiness floor cannot be evaluated.
• The extractor finds anchors when the engine sits outside the Protocol root.
• Executable examples reserve kind and unspecified for clear contract outcomes.
• The test wrapper routes Vitest options and combines split-run reports.
• The package declares Node versions that can load its CommonJS entry.
• Carriers read physical lines consistently, and repository Markdown links gain a check.
• The Studio labels gathered intents and lets people preview a memo before copying it.
• The repository’s review records and guidance describe the accepted inventory and updated rules.

No actionable defect was established. Tests and builds were not run during this review.

Acknowledged choices from darko-mijic were not reported again: path checks do not promise race-free confinement; inherited built-in methods remain eligible handlers; StepFn gains its third argument; subpaths remain ES-module-only; related always splits; blob reports and coverage remain refused. darko-mijic also explicitly leaves the temporal guard over docs/lineage/ and foreign stated readiness open.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Authored carriers and code anchors] --> B[Discover and reify]
  B --> C[Derive graph and validate]
  C --> D[Generate contracts]
  C --> E[Reader]
  E --> F[Render pages and notes]
  D --> G[Check output parent]
  F --> G
  G -->|Safe| H[Stage and publish]
  G -->|Unsafe| I[Refuse writes and cleanup through that parent]

Loading

_{Reviews (1) · Last reviewed commit: "fix: preserve cmark comment and referenc..." · Reviewed by Greptile}