feat(heroku-to-aws): make report validation blocking + harden it

[herosjourney]

Problem

Heroku migrations produce a migration-report.html — the summary a stakeholder reads. There's a checker that's supposed to make sure that report is complete and sound, but it wasn't actually enforcing anything: it was optional (a bad report didn't stop anything), it had no tests, and it only looked at a handful of the report's sections. On top of that, two of the skill's own instruction files disagreed about whether the check even blocks — one said "optional, never halt," the other said "fail the gate."

The practical result: a Heroku user could ship a report that's missing its required cost-optimization section, or shows a verdict signalled by color alone (a problem for colorblind readers), or has an empty section where content should be — and nothing would catch it. This PR makes the check actually block, hardens what it inspects, gives it a test suite, and resolves the contradiction so the two files tell one story.

In code, for reviewers

validate-heroku-migration-report.py was optional/non-blocking, had zero tests, and checked only three section IDs + the footer. generate-report.md said "optional / never-halt" while generate-assemble.md said "GATE_FAIL" — a direct contradiction. A report could ship missing cost-optimization, carrying a color-only verdict pill, or with an empty section, with no gate. Review item P1-B.

Solution

Harden the Heroku report validator, make it blocking, and add the test suite it never had —
scoped to what the intentionally-thin one-pager actually emits.

Validator (validate-heroku-migration-report.py), Python-stdlib only:

• Requires cost-optimization and enforces it is non-empty (a table or the explicit
  no-eligible-commitment sentence, never a blank section — matching the skill's Step 3 contract).
• Typography-first verdict: decision-summary must not contain any badge-verdict-*
  pill; when estimation-infra.json declares recommendation.outcome, it must render a
  verdict-headline.
• Accessibility (safe subset): <html lang>; every <th> declares scope=col|row; any
  <figure> carries aria-label + <figcaption>.

Emitter and gate now agree (a11y). generate-report.md was updated so the report template
instructs the a11y attributes the validator requires — every <th scope>, and figure labels if a
figure is emitted. Without this the blocking gate would have false-failed a report that followed
the template (the tables emit <th> with no scope). <html lang> was already in the skeleton.

Where the validator runs (round-2 fix — mirrors #287). Generate dispatches under
_exec._agent: rw, which excludes shell/Bash — the dispatched worker (fragments and
assembler) cannot run the validator. An earlier revision put the python3 invocation in
generate-assemble.md and called it "the main window," but the assembler is dispatched, so
"blocking" was prose on a path that can't execute. Corrected to the three-step split:

• Worker (assembler + generate-report.md, dispatched, no shell): authors
  migration-report.html and does the in-fragment Step 3 self-check + one-repair; it never runs
  the validator.
• "Finish Generate" (main window, BEFORE the gate): generate.md runs
  validate-heroku-migration-report.py ($PLUGIN_ROOT-anchored) over the report. On REPORT_OK
  it stamps a durable report-validation-status.json (temp status sidecar, not a _produces
  artifact); on REPORT_FAIL it emits GATE_FAIL and pastes the validator's errors[]
  verbatim so recovery is actionable. This step edits nothing else.
• _postconditions gate (main window, READ-ONLY): the section-id _assert plus a new
  _assert that report-validation-status.json has report_status == REPORT_OK. Runs no
  validator, edits nothing (per INTERPRETER.md § _postconditions). The durable stamp means the
  assert reads an artifact, not conversation memory — parity with feat(heroku-to-aws): wire tf-best-practices policy gate into Generate #287's validation-report.json.

Recovery on GATE_FAIL is a hand-edit of the report from the pasted errors + a direct
validator re-run (or a maintainer re-running Generate for a clean rebuild) — not a blind "re-run
Generate," which re-dispatches the shell-less worker and re-authors the report.

Blocking wiring resolved the contradiction. generate-report.md and generate-assemble.md
now agree the gate is required/blocking and read-only; generate.md _postconditions gained
cost-optimization in the section-id _assert plus the durable report-validation-status.json
REPORT_OK _assert.

Tests: tests/test_validate_heroku_migration_report.py (both trees, 14 cases) — compliant
report (with and without a declared recommendation) passes; failures fire for missing/empty
cost-optimization, missing footer, badge-verdict-* pill, <th> without scope, missing
<html lang>, missing verdict-headline when an outcome is declared, figure missing aria-label,
and (new) figure with aria-label but no <figcaption>; plus the figure-happy-path and
≥2-scenario paths.

Deliberately NOT ported (would false-fail the thin report): decision-before-TOC, single-<h1>,
per-table <caption>, glossary/appendix set.

Type of Change

• Bug fix
• Enhancement to existing content
• Guardrail/CI update

Team Folder

• advisor/
• migrate/
• solution-architecture/
• Other: ___

Verification

• cross-plugin-drift.ts: OK (273 identical, 25 allowlisted — no new drift)
• test_validate_heroku_migration_report.py (both trees): 13 passed each
• frontmatter-validator (both trees): OK
• markdownlint-cli2: 0 errors; dprint check: clean; tsc --noEmit: OK both trees
• False-fail probe: a template-shaped report (with <th scope>, non-empty cost-optimization) →
  exit 0; a bare <th> → exit 1 with the actionable message (gate matches emitter)
• mise run build (full composite incl. bandit/semgrep/gitleaks/checkov/grype):
  PASS, exit 0 — run against a fresh detached worktree at this branch's commit (not my working
  checkout, which has an untracked, unrelated migration-watchdog/ directory failing an unrelated
  markdownlint check). checkov 276/0; gitleaks no leaks; grype no vulnerabilities.

Not run locally: no browser a11y tooling (Vale/Pa11y/axe) — the validator is a dependency-free
static HTML reader, matching the plan's "Python 3 only" constraint.

Checklist

• I have read the CONTRIBUTING.md guidelines
• My changes do not include hardcoded secrets, credentials, or internal-only content
• I have run mise run build locally and it passes — verified against a fresh worktree at this branch's commit (see Verification)
• I have updated documentation if needed (generate-report.md a11y authoring rules + blocking wiring; generate-assemble.md main-window gate)
• My changes are scoped to my team's folder only (advisor/ + migrate/ migration skills)

[herosjourney]

Follow-up review + fixes (pushed as b8f4260)

A review pass over this PR found the blocking flip had real problems — a gate stricter than the
emitter, the same interpreter conflicts as the tf-policy-gate PR, and test gaps. All fixed.
Structured by what the review found and how it was addressed:

Blocking issues found

1. The a11y rules would false-fail a report that follows the template. The validator required
scope=col|row on every <th> and aria-label+<figcaption> on every <figure>, but
generate-report.md never told the agent to emit them — and the report emits tables
(exec-costs, cost-optimization, what-if-scenarios). An agent following the template would
produce <th>Tier</th> with no scope and the now-blocking validator would REPORT_FAIL. Same
"gate stricter than emitter" mistake I'd consciously avoided for <caption>.
→ Fix: added the a11y requirements to the emitter (generate-report.md Step 2 skeleton + the
Step 3 self-check): every <th> declares scope, any <figure> carries aria-label +
<figcaption>. Now the gate matches what the template produces — verified with a probe: a
template-shaped report passes, a bare <th> fails.

2. The completion gate said "GATE_FAIL … and repair," which the interpreter forbids.
INTERPRETER.md § _postconditions: a gate must not modify artifacts. And the paragraph above
the new blocking gate still said "repair the HTML and continue," reading as a second
continue-on-fail path.
→ Fix: the one-repair-then-rewrite lives only in the generate-report.md in-fragment Step 3
self-check; the assembler now emits GATE_FAIL and halts (tells the user to re-run Generate),
editing nothing. Step 3 was renamed and split into "in-fragment self-check (repair here)" vs
"blocking validation (main window)."

3. The required validator couldn't run under _exec: rw, and the path was cwd-relative. Same
host gap as the tf-policy-gate PR — the dispatched worker has no shell, so a "required, blocking"
validator invoked in the fragment is just prose. And generate-report.md used
python3 scripts/… (cwd-relative, misses the plugin script even on inline hosts).
→ Fix: the validator runs in the main window completion gate (generate-assemble.md +
generate.md _postconditions, never dispatched — INTERPRETER.md § _exec step 4), the only
place with a shell; the worker only writes the HTML. Path corrected to
"$PLUGIN_ROOT/scripts/validate-heroku-migration-report.py".

Should-fix issues found

4. Docstring vs code on badge-verdict-*. The docstring said "as the sole verdict carrier" but
the code (correctly) fails on any badge-verdict-*. Three sentences disagreed.
→ Fix: aligned docstring, code comment, and generate-report.md Step 3 to "must not contain
badge-verdict-*"; Step 3 now also requires a verdict-headline when recommendation.outcome is
set, matching the validator.

5. Test gaps for rules the PR made blocking. No figure-a11y test, no happy-path test with a
declared recommendation.outcome, and empty cost-optimization passed (only the id was checked).
→ Fix: the validator now enforces non-empty cost-optimization; added tests for
figure-without-aria-label (fails), figure-with-label (passes), good-report-with-recommendation
(passes), and empty-cost-optimization (fails). 9 → 13 cases. Renamed Step 3 from "Soft gate."

6. Corrupt estimation-infra.json fails open silently.
→ Fix: added a one-line comment documenting the intentional fail-open-on-ambiguity.

Verification (b8f4260, fresh detached worktree)

mise run build PASS (checkov 276/0, gitleaks no leaks, grype no vulns), cross-plugin-drift.ts
OK (273 identical), test_validate_heroku_migration_report.py 13 passed both trees, all touched
files twin-identical, false-fail probe confirms the gate matches the emitter. Not run locally:
browser a11y tooling — the validator is a dependency-free static reader.

[herosjourney]

Verified additional issues and each point against the code at b8f4260 before changing anything — all four hold, and #1 is the same host-gap as the first #287 pass, one file over. Fixed in 2868f78 (both trees, byte-identical). Structured as finding → what I changed.

1. Should-fix — validator was still in the worker, not the main window (agreed, blocking)

Confirmed: the python3 validate-heroku-migration-report.py call lived only in generate-report.md (fragment) and generate-assemble.md (assembler) — both dispatched to the shell-less rw worker. generate.md _postconditions only asserted section-ids + footer prose; it never ran the script or asserted REPORT_OK. So on _exec: rw the new checks (no badge-verdict-*, verdict-headline, <th scope>, non-empty cost-optimization, <html lang>) never executed, and a report with the four section-ids + footer passed the only main-window check. The assembler even labelled itself "the main window," which is wrong — the assembler is dispatched. Mirrored the #287 split exactly:

• generate.md — new main-window "Finish Generate in the main window (report validation)" step that runs validate-heroku-migration-report.py after the worker returns and before _postconditions. Added a read-only _assert that this step printed REPORT_OK (not just that the section-ids exist). Scope Boundary carves the main-window validate out of "Nothing else."
• _postconditions is read-only: the section-id _assert + the REPORT_OK _assert, no script run, no edits.
• generate-assemble.md — dropped the false "In the main window … run the report validator" claim; it now states the assembler is dispatched and points at the finish step; halts read-only on GATE_FAIL.
• generate-report.md — points at the finish step for authoritative execution.

2. Nit — GATE_FAIL should quote the validator errors (agreed)

Right — the fragment has no shell to re-run the validator, so a bare "re-run Generate" is not actionable. The finish step now emits GATE_FAIL and pastes the validator's errors[] verbatim (missing scope id, empty cost-optimization, banned pill class, <th> without scope, missing <html lang>, etc.). I also fixed the recovery copy the same way #287 was fixed: recovery is a hand-edit from the pasted errors + a direct validator re-run, not a full Generate re-dispatch (which re-authors the report under the shell-less worker and wipes the fix).

3. Nit — skeleton had no <th scope="col"> example (agreed)

The HTML skeleton showed <section id="exec-costs">…</section>. Agents copy skeletons, so <html lang="en"> got followed but <th scope> might not. The exec-costs skeleton section now shows a sample table with <th scope="col"> header cells and a comment to copy the pattern for cost-optimization / what-if-scenarios.

4. Nit — no test for figure-without-<figcaption> (agreed)

Added test_figure_without_figcaption_fails — a <figure aria-label="…"> with no <figcaption> now asserts exit 1 + a figcaption error, symmetric to the existing missing-aria-label case. Test count 13 → 14 (both trees).

Verification (2868f78, both trees)

• cross-plugin-drift.ts: OK (273 identical, 25 allowlisted)
• test_validate_heroku_migration_report.py: 14 passed advisor, 14 passed migrate
• mise run build against a fresh detached worktree at this commit: PASS, exit 0 (checkov 276/0, gitleaks no leaks, grype no vulnerabilities, fail 0 across DSL suites)
• Fail-closed preserved: if the finish step can't run the validator (no shell), it does not record REPORT_OK, so the read-only _assert fails closed.

No change to the a11y/emitter alignment, the assembler-no-longer-edits-HTML property, the $PLUGIN_ROOT path, the badge-verdict-* wording, or the not-ported set — those stay as you approved.

[herosjourney]

Fixed 3 additional nits.

1. REPORT_OK was conversation-only (agreed — added parity)

Correct: #287 stamps a durable policy_status into validation-report.json, but #288's finish step only "recorded" a print and the _assert read prose. Since the validator only emits to stdout/stderr, I added the one-line stamp you suggested: the Finish Generate step now writes report-validation-status.json ({"report_status": "REPORT_OK"}, or REPORT_FAIL with errors[]), and the read-only _postconditions _assert now checks that file's report_status == REPORT_OK. So the assert reads an artifact, not conversation memory — parity with #287. It's a temp status sidecar (not in _produces), and if the validator can't run the stamp is absent/not_run, so it still fails closed. generate-report.md and generate-assemble.md reference the stamp consistently, and recovery now says re-run + re-stamp.

2. Assembler header was stale (agreed)

Right — the opener still said the assembler "enforces the completion handoff gate." Rewrote it: the assembler runs the cross-artifact checks and emits the handoff signal, but the report validator runs later in the main-window "Finish Generate" step (the assembler is dispatched and has no shell). Body was already correct; this aligns the opener.

3. Two GATE_FAIL emitters (agreed it's redundant — leaving it)

Agreed it's fine and fail-closed. The redundancy is deliberate defense-in-depth: the finish step emits GATE_FAIL on REPORT_FAIL, and the read-only _postconditions _assert independently blocks on a non-REPORT_OK stamp — so a skipped/half-run finish step still can't complete. I'd rather keep both than rely on either alone, so no change here.

Verification (2bce3f5, both trees)

• cross-plugin-drift.ts: OK (273 identical, 25 allowlisted)
• test_validate_heroku_migration_report.py: 14 passed (both trees)
• mise run build against a fresh detached worktree at this commit: PASS, exit 0 (checkov 276/0, gitleaks no leaks, grype no vulnerabilities)

No validator-code or test change in this revision — it's the durable-stamp wiring + the assembler-opener wording.

[leon1418]

[🤖 AI review 🤖]

Reviewed head d35f4b8bb65d673d541ed03ab9f7f0346abc32d0, including the follow-up explanations and both plugin copies. Requesting changes for the three inline findings:

• P1: The newly mandatory validator is not included in the documented skill-only installation.
• P2: Heading-only cost-optimization sections and tables without data rows still pass.
• P2: The blocking accessibility checks reject valid HTML attribute syntax.

Validation: all 28 added tests passed across both trees; all five changed mirror pairs are byte-identical; cross-plugin drift and both Heroku frontmatter checks passed; current GitHub checks are green. Additional probes reproduced the validator gaps in both copies, and a skill-directory packaging probe confirmed the missing script. These cases are not covered by the added tests.

[herosjourney]

Thanks — reproduced all three before replying:

1. Validator doesn't ship with a standalone skill install. Checked a real --skill heroku-to-aws install on disk: no scripts/ directory, no validator file anywhere under it, and $PLUGIN_ROOT is never defined inside the skill. This is a real blocker for that documented install path, not a hypothetical.
2. Cost-optimization content check accepts decorative-only markup. Confirmed through the full CLI: a heading-only section and a table with headers but an empty <tbody> both pass REPORT_OK; a control case with an actual opportunity row behaves correctly once the section's other structural rules are met.
3. Attribute regexes reject valid HTML syntax. Confirmed: scope="col" passes, scope = "col" and scope=col (both valid, equivalent HTML) fail the now-blocking validator.

All three will be fixed in both trees: bundle the validator inside the skill and resolve its path from the skill root, exclude headings/table headers from the content check and require real rows or the explicit no-eligible statement, and parse attributes properly instead of matching literal syntax. Replied inline on each with the specific fix.

[leon1418]

[🤖 AI review 🤖]

Re-reviewed exact head 8ddd56436b48e1c5d166e9f6bb1d7d150b916188 against base 0f067dfe8b4da3f9cca969f1c2022a75b4b7f154, including all author replies, the full PR diff, both plugin trees, and the Generate worker/main-window handoff.

The standalone validator packaging P1 is fixed. I replied in all three original threads without creating duplicate findings. Two P2 concerns remain:

• A table with no opportunity rows still passes when accompanied by the credits disclaimer that the report template requires.
• The accessibility attributes are normalized, but verdict class attributes still produce false failures for valid syntax and miss an unquoted banned badge.

Local validation: 46/46 tests passed; isolated skill-only validator execution passed for both copies; all six changed surviving mirror pairs are byte-identical and the old script is removed in both trees; cross-plugin drift (277 identical, 25 allowlisted), both Heroku frontmatter checks, both vendored-shared checks, and git diff --check passed. I also inspected the Heroku allowlisted interpreter/schema differences and ran 40 targeted CLI probes; the remaining failures are reproduced in both copies. No full agent-driven migration replay was run.

Current GitHub CI is green, separately from those local results. Retaining Request changes until the two remaining validator gaps are addressed.

[herosjourney]

Follow-up fixes (pushed as 4dddc4a1)

Both remaining P2s from the last review round are fixed in both plugin trees:

1. Cost-optimization content check accepted decorative-only markup. _validate_optimization_content() checked for "any text left after stripping headings/<th>", so an empty <tbody> paired with the credits disclaimer (or any other prose) satisfied it. Replaced with a positive check: pass only on a real <tbody> row with non-empty <td> text, or the exact no-eligible-commitment sentence. Nothing else counts. Replied inline with the specific repro coverage.

2. Verdict checks weren't using tokenized class attributes. _validate_verdict() still matched class="..." literally, unlike the scope/lang/aria-label checks that already went through HTMLParser. Added _class_tokens() (same _collect_tags() machinery) and switched both the badge-verdict-* ban and the verdict-headline requirement to set-membership checks. Replied inline with the specific repro coverage.

Also merged current main (this PR's branch had fallen behind — main since gained the rwx capability tier + Terraform policy gate on this same phase, and the resulting generate.md conflict was resolved by combining both main-window exceptions in the Scope Boundary paragraph: the policy checker's scoped shell during assembly, and the report validator's separate main-window "Finish Generate" step).

Verification (4dddc4a1, both trees)

• test_validate_heroku_migration_report.py: 29 passed (14 → 29; added 6 regression tests for the two P2s plus retained coverage)
• cross-plugin-drift.ts: OK (283 identical, 27 allowlisted)
• mise run lint:frontmatter: OK (both trees)
• mise run fmt:check (dprint): clean
• mise run lint:md (markdownlint-cli2): 0 errors across 896 files
• Manually reproduced both reviewer repro cases against the pre-fix validator to confirm they failed as described, then confirmed the fix resolves them

@leon1418 please re-review the current head when convenient.

[leon1418]

[🤖 AI review 🤖]

Re-reviewed exact head 4dddc4a14098ae1e4eb71a46cd8342215e2280ce against base b30cb39be02d2eae78cb8c482f4018435d4dc988, including the author's latest replies, the current diff, both plugin copies, and the merged rwx policy-checker/main-window report-validation integration.

The standalone packaging P1 remains fixed. The verdict-class P2 is now fixed, and the earlier credits-disclaimer repro is fixed. I replied in the two existing P2 threads. One P2 content-validation concern remains: the new regex row check accepts blank character-reference cells and commented-out rows, while rejecting a populated table whose source omits optional tbody tags. The original content thread contains the reproductions and suggested correction; no duplicate thread was created.

Local validation: 58/58 tests passed; all eight before/after checks reproduced the earlier bugs and verified the claimed corrections across both mirrors; isolated skill-only validator execution passed in both copies; 64 scoped CLI probes covered original and additional cases. Cross-plugin drift (283 identical, 27 allowlisted), both Heroku frontmatter checks, both vendored-shared checks, and git diff --check passed. The changed validator/tests/Generate files match across mirrors; the allowlisted SKILL.md difference is the existing advisor-only offers block. No full agent-driven migration replay or deployment was run.

Current GitHub CI is green, separately from those local checks. Retaining Request changes for the remaining content-validation gap.

[leon1418]

[🤖 AI review 🤖]

Re-reviewed exact head b62d8c67aeddb7c9fa2fd36e7102a60ca71cb8ab against base ade57adaf9d776d18d316d28d3cee3f746582ced, including the author reply, current diff, both mirrors, and Generate's report-validator integration.

The latest entity-only, commented-row, and implicit-tbody cases are fixed. Packaging, substantive-content controls, and parsed verdict classes retain their earlier fixes. One P2 remains: a populated table with a valid omitted td end tag is rejected because the new parser only commits cell content on an explicit </td>. I added the reproduction and suggested correction to the original content thread; no duplicate thread was created.

Local validation: 66/66 tests passed; eight before/after comparisons reproduced and verified the newly claimed fixes in both copies; isolated skill-only validator execution passed in both copies. Of 64 scoped CLI probes, only the omitted-td-ending case remains incorrect (once per mirror). Cross-plugin drift (283 identical, 27 allowlisted), both Heroku frontmatter checks, both vendored-shared checks, and git diff --check passed. I checked the intentional Heroku SKILL.md/interpreter/schema differences and confirmed the main-window report-validation placement. No full agent-driven migration replay or deployment was run.

Current GitHub CI is green, separately from local validation. Recommendation: Request changes for the remaining parser boundary case.

[leon1418]

[🤖 AI review 🤖]

Completed the follow-up review of exact head e3ca4d232e9592c742e40dcaf30903f0e876ca90 against base ade57adaf9d776d18d316d28d3cee3f746582ced. The optional-td-end finding is fixed, and all previously reported packaging, content-validation, and attribute/class findings remain addressed. I found no remaining substantive concerns and replied in the original content thread.

Validation: 72/72 tests passed across both trees; 92/92 scoped CLI probes matched their expected outcomes; ten before/after comparisons reproduced the prior failures and verified the corrected behavior and negative controls. Isolated skill-only validator execution passed for both copies. Cross-plugin drift (283 identical, 27 allowlisted), both Heroku frontmatter checks, both vendored-shared checks, and git diff --check passed. The intentional SKILL.md/interpreter/schema differences and the main-window report-validation integration were checked. No full agent-driven migration replay or deployment was run.

Current GitHub CI is green, separately from local validation. Recommendation: Approve, for the user to submit. This is a COMMENT follow-up only; I have not submitted approval, dismissed prior reviews, or resolved conversations. The existing blocking review state still requires the user's action.

[leon1418]

[🤖 AI review 🤖]

Reviewed exact head 2ecdcfcd06c10b0a3ce5e2fd6f6882bc7256e5b9 against base 811fddafced0870173dc9c1a66749a9143baa040, including the currency-validation merge, relocated tests, standalone skill layout, both plugin copies, and the current discussions. The previously reported packaging, content, optional-td-ending, and attribute/class issues remain fixed.

One integration follow-up is inline: the now-mandatory currency check requires whole-dollar monthly values, but the dispatched Heroku report writer and its self-check do not carry that rule. Please align the producer instructions with the gate. Recommendation: Request changes for this P2 instruction/validator mismatch.

Validation: 158 focused tests passed; all 122 Heroku CLI probes and four isolated skill-packaging checks passed. Of 74 GCP currency probes, two expose an implicit-cell-boundary behavior already present in the byte-identical merged-base GCP validator; I am not attributing it to this PR. The relocated Heroku currency helpers match main, and the prior structural/verdict helpers are unchanged. Cross-plugin drift (284 identical, 27 allowlisted), both Heroku frontmatter checks, both vendored-shared checks, and git diff --check passed. GitHub CI is green separately from local validation. No full agent-driven migration replay or deployment was performed.

This is a COMMENT review; the existing human approval and historical reviews are left untouched.

[leon1418]

[🤖 AI review 🤖]

Completed startups-hybrid-v1 review of exact head 571f7daaaa419e221e6d4c9fa5404ea9d5c19778 against base ec681ba5d49b039dbf1ef9d87d3304205b824f3d: actual OCR v1.12.5 delegation, one fresh general reviewer, the parent repository review, and independent candidate verification. All 18 Git inventory entries were covered, including Markdown and deletions. The canonical evidence gate passed.

Two verified P2 concerns remain, updated in the original threads:

1. The dispatched report writer still lacks the monthly rounding instruction required by the blocking currency gate.
2. The no-commitment fallback rejects valid encoded text and can accept content present only in comments/inert templates.

Local validation: the 108 authored tests passed in both independent perspectives; the parent ran 34 CLI/packaging checks, the general reviewer ran 48 focused base/head CLI probes and four packaging checks, and the parent independently ran 14 fallback verification cases. Drift, frontmatter, vendored-source, and diff checks passed. Current GitHub CI is green separately from those results. No full agent-driven migration replay, browser-accessibility run, or deployment is claimed.

Recommendation: Request changes for the two P2s. No approval, dismissal, merge, or conversation-resolution action was performed.

[leon1418]

[🤖 AI review 🤖]

Completed startups-hybrid-v1 at exact head e538f368cb388fa36e6dbbb0c4e8cd52c58650d2 against base ec681ba5d49b039dbf1ef9d87d3304205b824f3d: OCR v1.12.5 delegation, one fresh general review, parent repository review, independent candidate verification, and the canonical gate. All 18 Git entries were covered. The actual child's initial-context disclosure and recorded tool audit passed the isolation checks.

The currency-authoring and decoded-fallback fixes are verified. Two P2 concerns remain: inert-only opportunity rows can satisfy the content check, and an inert-only verdict headline can satisfy the declared-outcome check. Both were independently reproduced with supported Estimate inputs and updated in the original threads. The base lacked these positive-content requirements; this is incomplete enforcement of the newly introduced guarantees.

Validation: 116 head tests passed in both perspectives; the general reviewer also ran 36 base currency tests. The parent ran 54 CLI/packaging checks and 26 authoritative candidate-verification executions. Drift, frontmatter, vendored-source, and diff checks passed. GitHub CI is green separately from local results. No full agent-driven migration replay, browser-accessibility run, or deployment was performed.

Recommendation: Request changes for the inert-content cases. No approval, dismissal, merge, or conversation-resolution action was performed.

[leon1418]

[🤖 AI review 🤖]

Completed startups-hybrid-v1 with convergence v1 at exact head ab95346fcad2bc3f432489d09a7c1696a5b9f7e3, base ec681ba5d49b039dbf1ef9d87d3304205b824f3d: OCR v1.12.5 delegation, one fresh general reviewer, separate parent review and impact maps, candidate verification, and the canonical evidence gate. All 18 Git entries were covered.

The two previously reported inner-inert cases are fixed. Packaging, currency guidance, decoded fallback text, optional cell endings and class handling retain their fixes. Two consolidated P2 corrections remain:

1. Section parsing: in both skill-local validators and structural test copies, use actual parsed section IDs with document/comment/inert context. Invisible sections and data-id must not satisfy the gate; real quoted/spaced/unquoted IDs must pass; commented duplicates and the what-if skeleton comment must not count. The original content and verdict threads carry the reproductions.
2. Missing-Python instructions: update the tooling fallback in both SKILL.md copies so it no longer promises Generate completion without successful report validation. Earlier phases can continue; the existing missing/not_run gate stays blocking. This is inline.

Both gaps already existed at e538f368; I missed them in the prior review. Neither is attributed to the latest inert-parser fix.

Local evidence: 124 authored tests passed in each perspective; the parent ran 54 CLI/packaging checks and 76 authoritative before/after boundary executions with preserved inputs. The fresh reviewer separately reproduced the section issue. Drift, frontmatter, vendored-source and diff checks passed. Current GitHub CI reports nine successful checks, separately from local results. No full agent-driven migration, missing-Python host replay, browser accessibility suite or deployment was performed.

Recommendation: Request changes. GitHub also reports merge conflicts, a separate landing blocker. Approval remains for the user; no approval or review dismissal was performed.