[branben]

What this is

Bead 1nzr — reorganise the flat root into src/{bridge,core,crew,review,sensors,tools}. Commits:

commit	what
b372f7b	the reorg itself (93 modules, reorg-only — no behaviour change)
c3bfd78	repair the imports the cherry-pick reverted to flat paths
8e826e2	stable ruff F-class gate + clean the newly-covered tree
f6705de	follow the tests' flat imports and patch targets to the new layout
96fe03c	index-gate selftest follows verify_gate.py to src/review/
a2afda4	anchor root-relative paths after the move; hermetic adversarial-review test
bb1b968	move the bundled behaviour changes out to #153 (keep this PR reorg-only)

Behaviour changes moved out

Earlier revisions of this PR also carried behaviour changes. Commit bb1b968 removed them, and they now live in the stacked PR #153 (#153):
onboard.py and the --onboard CLI, a2a_bridge.py, the ENTIRE CRITICAL/HIGH veto on the sync path, the conductor bookbag persistence of entire findings, the entire_review rename and CLI change (--profile general), the review_packet._sensor list/int fix, and the docs describing them. This PR is now main's behaviour plus the move.

Note: .github/workflows/ci.yml still lists onboard.py a2a_bridge.py in the ruff CORE list. Those files no longer exist on this branch, so that line needs a separate push with workflow scope.

Measured

python3 -m pytest tests/ -q --tb=line --basetemp=/tmp/pytest-reorg-final
  4 failed, 2474 passed, 34 skipped, 7 warnings in 740.19s

This matches the bead's recorded baseline exactly on passed/skipped (2474P / 34S), with one fewer failure. The bead recorded 5F/2474P/34S; this branch measures 4F. I am treating my own measurement as the baseline rather than quoting the bead's — the discrepancy is that the bead's 5th failure was a test_issue_bridge case this branch fixes.

All 4 remaining failures are pre-existing and environmental, not introduced here:

failure	cause
test_omni_relay_live ×3	src/tools/omni_relay.py:364 → TransportError: upstream HTTP 500
test_timeout_marker	pytest plugin config; passes in isolation (order-dependent)

What the reorg broke, and how it was found

b372f7b moved modules without following the tests' references to them. 16 failures:

file	was	now
tests/test_leaf.py	5 failed	42 passed
tests/test_anchor_loader.py	6 failed	66 passed
tests/test_issue_bridge.py	5 failed	158 passed

Two distinct traps, both now written into the codebase-restructuring skill:

1. import src.core.leaf binds src, not leaf. The first rewrite of test_anchor_loader.py produced NameError: name 'prompt_composer' is not defined instead of fixing anything — the alias must survive (import src.tools.prompt_composer as prompt_composer).
2. patch("module.fn") is a STRING target, so an import sweep cannot see it. It fails only at test-run time while collection stays green.

Sweep coverage: a whole-repo import sweep found zero flat imports of moved modules outside .scratch/. A whole-repo sweep for string references found 72 sites, of which 67 were false positives — evidence strings ("crew_dispatch.py:123"), argv (["teacher.py", …]), filenames ("leaderboard.html"), docstrings, and a vendored repo's own trajectory.json. Each was checked individually rather than bulk-rewritten.

Lint gate

The lint step was scoped to "core modules + any .py changed in this PR". A diff-scoped gate assumes a changed file is a file someone just wrote; the reorg rewrote intra-repo imports across the repo, so ~280 files were "changed" while only their import lines moved. The gate is now scoped to the canonical tree.

review-gate is the only required check — the lint step lives in the non-required test job.

python3 -m ruff check issue_bridge.py conductor.py review_packet.py bookbag.py src/ scripts/
  → All checks passed!        (was 107 errors)
python3 -m ruff check --select F821,F811,F822 tests/
  → All checks passed!        (was 2 errors)

Real defects the wider scope exposed: a phantom __all__ export in src/core/persistence_contract.py:466 (AUDIT_IDENTITY_FIELDS — the real name is _AUDIT_IDENTITY_FIELDS, so import * would raise AttributeError), and scripts/candidate_pr_contract.py calling CandidateStore/SimpleNamespace with no import.

Deliberately NOT deleted

Two lint findings are missing features, not stale code. Deleting them would satisfy the linter while cementing the gap — both are filed as beads:

• school-core-tkqn — generate_activity_dashboard.py builds data = json.dumps(...) and never uses it, while the docstring promises "all data baked in".
• school-core-88jf — bookbag.py constructs BookbagSignal(bead, repo) and never reads it; the polling loop still reads the bookbag JSON, so the flag-file fast path is dead.

Both are kept, # noqa: F841, with the reason in the comment.

Verification status

• Full suite with a real summary line, above.
• ruff green on the exact CI commands.
• Honest import probe: 109 modules, 109 OK, 0 FAIL.
• 45 changed files compile.

What this does NOT fix

• No runtime behaviour was changed — every edit is an import, a dead-statement removal, or a comment.
• The two noqa'd findings above remain open beads.

Confidence Score: 0/5

Do not merge yet: scheduled execution breaks, verification paths are wrong, and the new review veto can still allow publication.

Findings

1.  Local reviews check wrong folder ▶
2.  Critical findings still allow publication ▶
3.  Lock timeout drops the veto ▶
4.  New dependencies cannot be locked ▶
5.  Dependency pins cross repositories ▶
6.  Explicit dependency pins disappear ▶
7.  Linked code differs from lock ▶
8.  Broken config passes verification ▶
9.  Evidence requirement has no effect ▶

Fix with agent prompt

### Issue 1
src/core/director.py:1861-1863
The current-checkout fallback in `_resolve_repo_path` now returns `src/core`, not the checkout root. A local task for the current repository with no cached clone therefore freezes and verifies that subtree, missing the root `project_verify.yaml` and other source files. Return the checkout root here.

```suggestion
        from src.tools.repo_default import default_repo
        if repo == default_repo():
            return Path(__file__).resolve().parent.parent.parent
```

### Issue 2
issue_bridge.py:2713-2714
The new Entire veto only changes `adversarial_review`. It does not reject `canonical_packet` or stop the loop. With passing build checks and judges, a CRITICAL sensor finding still reaches grading and PR publication, while the saved packet remains accepted. Reject the authoritative result and stop before those side effects.

### Issue 3
conductor.py:657-664
The new persistence step ignores `locked_update_bookbag` returning `None`, its documented lock-timeout result. If this write times out and the lock later clears, `_persist_acceptance` can read no Entire findings and accept both passing judges despite a CRITICAL sensor result. Check the return value and prevent acceptance when the result was not saved.

### Issue 4
onboard.py:1310-1313
`_infer_dependencies` stores sources as `owner/repo`, but this URL guard only expands strings containing `github.com/`. Collected sources therefore never reach the online lookup. A new dependency using a branch or tag gets `commit: null` unless an existing lock supplies it, even online, and fails the documented lock check. Normalize the slug into a clone URL and use the slug for the commit lookup.

### Issue 5
onboard.py:1301-1306
Existing lock entries are reused by basename alone, without checking the owner or requested ref. A target's `acme/ripwire` dependency matches the orchestrator's `redhat-et/ripwire` entry and copies that other repository's source and commit. A changed ref for the same repository also reuses the old pin. Match the full normalized source and requested ref before reusing a lock.

### Issue 6
onboard.py:691-692
The raw-text scan adds each GitHub dependency with `ref="HEAD"` before the parsed `[tool.uv.sources]` entry supplies its `rev` or `tag`. Because `_add` uses `setdefault`, that explicit pin is discarded. Parse structured entries first, or allow an explicit pin to replace the guessed `HEAD` value.

### Issue 7
onboard.py:1547-1549
`_envit_store_path` returns the first checkout directory without comparing it with the generated lock's commit. If the store holds another revision, the new `.envit/repos` link points there while the lock claims a different pin. Existing links are also skipped on refresh. Select and verify the locked commit, and refresh links when the pin changes.

### Issue 8
onboard.py:183-191
The generated YAML and JSON checks end with `|| true`, so malformed files still produce a successful exit. `run_verify_gate` treats that exit as a passed check. A newly onboarded config-only repository can therefore report successful verification without any check capable of failing. Preserve parser failures instead of forcing success.

```suggestion
        {"name": "yaml-validate",
         "cmd": "python3 -c \"import yaml, sys; [yaml.safe_load(open(f)) for f in sys.argv[1:]]\" "
                "$(find . -name '*.yaml' -o -name '*.yml' | grep -v node_modules | head -5) 2>&1",
         "cwd": ".", "languages": "yaml"},
    ],
    "json": [
        {"name": "json-validate",
         "cmd": "python3 -c \"import json, sys; [json.load(open(f)) for f in sys.argv[1:]]\" "
                "$(find . -name '*.json' | grep -v node_modules | head -5) 2>&1",
```

### Issue 9
onboard.py:821
`_collect_evidence` always adds `package_manager` to `generated`, even when it only chose the default `npm` and read no repository facts. Since `detected` is just `bool(generated)`, `--require-evidence` cannot reject an empty initialized checkout. Only mark facts actually read from the target, and test this with a real empty Git checkout.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Moves root modules into six src packages and updates imports, tests, and lint scope. The diff also adds onboarding and A2A tools and changes Entire review gating.

• The project’s Python modules now live in six named source groups.
• Repository onboarding can turn project evidence into setup files.
• The A2A bridge sends tasks to OpenHands and retrieves their answers.
• Entire review findings now feed into issue acceptance.

No runtime tests were executed during this review.

Acknowledgments from branben: the dashboard's unused payload and the bookbag flag-file fast path are intentionally deferred as school-core-tkqn and school-core-88jf. The PR description also identifies three live-relay failures, the order-dependent timeout failure, and the full-suite malformed-JSON test failure as known or environmental; none is reported here.

_{Reviews (1) · Last reviewed commit: "fix(reorg): follow the tests' flat impor..." · Reviewed by Greptile}

[sourcery-ai]

Sorry @branben, your pull request is larger than the review limit of 150,000 diff characters

[qodo-code-review]

ⓘ Qodo reviews are paused because the subscription is no longer active. Ask your workspace admin to reactivate the subscription to resume reviews. Manage billing

[greptile-apps]

Local reviews check wrong folder

The current-checkout fallback in _resolve_repo_path now returns src/core, not the checkout root. A local task for the current repository with no cached clone therefore freezes and verifies that subtree, missing the root project_verify.yaml and other source files. Return the checkout root here.

Suggested change

	from src.tools.repo_default import default_repo
	if repo == default_repo():
	return Path(__file__).resolve().parent
	from src.tools.repo_default import default_repo
	if repo == default_repo():
	return Path(__file__).resolve().parent.parent.parent

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/core/director.py
Line: 1861-1863

Comment:
**Local reviews check wrong folder**

The current-checkout fallback in `_resolve_repo_path` now returns `src/core`, not the checkout root. A local task for the current repository with no cached clone therefore freezes and verifies that subtree, missing the root `project_verify.yaml` and other source files. Return the checkout root here.

```suggestion
        from src.tools.repo_default import default_repo
        if repo == default_repo():
            return Path(__file__).resolve().parent.parent.parent
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Critical findings still allow publication

The new Entire veto only changes adversarial_review. It does not reject canonical_packet or stop the loop. With passing build checks and judges, a CRITICAL sensor finding still reaches grading and PR publication, while the saved packet remains accepted. Reject the authoritative result and stop before those side effects.

Prompt To Fix With AI

This is a comment left during a code review.
Path: issue_bridge.py
Line: 2713-2714

Comment:
**Critical findings still allow publication**

The new Entire veto only changes `adversarial_review`. It does not reject `canonical_packet` or stop the loop. With passing build checks and judges, a CRITICAL sensor finding still reaches grading and PR publication, while the saved packet remains accepted. Reject the authoritative result and stop before those side effects.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Lock timeout drops the veto

The new persistence step ignores locked_update_bookbag returning None, its documented lock-timeout result. If this write times out and the lock later clears, _persist_acceptance can read no Entire findings and accept both passing judges despite a CRITICAL sensor result. Check the return value and prevent acceptance when the result was not saved.

Prompt To Fix With AI

This is a comment left during a code review.
Path: conductor.py
Line: 657-664

Comment:
**Lock timeout drops the veto**

The new persistence step ignores `locked_update_bookbag` returning `None`, its documented lock-timeout result. If this write times out and the lock later clears, `_persist_acceptance` can read no Entire findings and accept both passing judges despite a CRITICAL sensor result. Check the return value and prevent acceptance when the result was not saved.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

New dependencies cannot be locked

_infer_dependencies stores sources as owner/repo, but this URL guard only expands strings containing github.com/. Collected sources therefore never reach the online lookup. A new dependency using a branch or tag gets commit: null unless an existing lock supplies it, even online, and fails the documented lock check. Normalize the slug into a clone URL and use the slug for the commit lookup.

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 1310-1313

Comment:
**New dependencies cannot be locked**

`_infer_dependencies` stores sources as `owner/repo`, but this URL guard only expands strings containing `github.com/`. Collected sources therefore never reach the online lookup. A new dependency using a branch or tag gets `commit: null` unless an existing lock supplies it, even online, and fails the documented lock check. Normalize the slug into a clone URL and use the slug for the commit lookup.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Dependency pins cross repositories

Existing lock entries are reused by basename alone, without checking the owner or requested ref. A target's acme/ripwire dependency matches the orchestrator's redhat-et/ripwire entry and copies that other repository's source and commit. A changed ref for the same repository also reuses the old pin. Match the full normalized source and requested ref before reusing a lock.

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 1301-1306

Comment:
**Dependency pins cross repositories**

Existing lock entries are reused by basename alone, without checking the owner or requested ref. A target's `acme/ripwire` dependency matches the orchestrator's `redhat-et/ripwire` entry and copies that other repository's source and commit. A changed ref for the same repository also reuses the old pin. Match the full normalized source and requested ref before reusing a lock.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Explicit dependency pins disappear

The raw-text scan adds each GitHub dependency with ref="HEAD" before the parsed [tool.uv.sources] entry supplies its rev or tag. Because _add uses setdefault, that explicit pin is discarded. Parse structured entries first, or allow an explicit pin to replace the guessed HEAD value.

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 691-692

Comment:
**Explicit dependency pins disappear**

The raw-text scan adds each GitHub dependency with `ref="HEAD"` before the parsed `[tool.uv.sources]` entry supplies its `rev` or `tag`. Because `_add` uses `setdefault`, that explicit pin is discarded. Parse structured entries first, or allow an explicit pin to replace the guessed `HEAD` value.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Linked code differs from lock

_envit_store_path returns the first checkout directory without comparing it with the generated lock's commit. If the store holds another revision, the new .envit/repos link points there while the lock claims a different pin. Existing links are also skipped on refresh. Select and verify the locked commit, and refresh links when the pin changes.

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 1547-1549

Comment:
**Linked code differs from lock**

`_envit_store_path` returns the first checkout directory without comparing it with the generated lock's commit. If the store holds another revision, the new `.envit/repos` link points there while the lock claims a different pin. Existing links are also skipped on refresh. Select and verify the locked commit, and refresh links when the pin changes.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Evidence requirement has no effect

_collect_evidence always adds package_manager to generated, even when it only chose the default npm and read no repository facts. Since detected is just bool(generated), --require-evidence cannot reject an empty initialized checkout. Only mark facts actually read from the target, and test this with a real empty Git checkout.

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 821

Comment:
**Evidence requirement has no effect**

`_collect_evidence` always adds `package_manager` to `generated`, even when it only chose the default `npm` and read no repository facts. Since `detected` is just `bool(generated)`, `--require-evidence` cannot reject an empty initialized checkout. Only mark facts actually read from the target, and test this with a real empty Git checkout.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Broken config passes verification

The generated YAML and JSON checks end with || true, so malformed files still produce a successful exit. run_verify_gate treats that exit as a passed check. A newly onboarded config-only repository can therefore report successful verification without any check capable of failing. Preserve parser failures instead of forcing success.

Suggested change

	{"name": "yaml-validate",
	"cmd": "python3 -c \"import yaml, sys; [yaml.safe_load(open(f)) for f in sys.argv[1:]]\" "
	"$(find . -name '*.yaml' -o -name '*.yml' | grep -v node_modules | head -5) 2>&1 || true",
	"cwd": ".", "languages": "yaml"},
	],
	"json": [
	{"name": "json-validate",
	"cmd": "python3 -c \"import json, sys; [json.load(open(f)) for f in sys.argv[1:]]\" "
	"$(find . -name '*.json' | grep -v node_modules | head -5) 2>&1 || true",
	{"name": "yaml-validate",
	"cmd": "python3 -c \"import yaml, sys; [yaml.safe_load(open(f)) for f in sys.argv[1:]]\" "
	"$(find . -name '*.yaml' -o -name '*.yml' | grep -v node_modules | head -5) 2>&1",
	"cwd": ".", "languages": "yaml"},
	],
	"json": [
	{"name": "json-validate",
	"cmd": "python3 -c \"import json, sys; [json.load(open(f)) for f in sys.argv[1:]]\" "
	"$(find . -name '*.json' | grep -v node_modules | head -5) 2>&1",

Prompt To Fix With AI

This is a comment left during a code review.
Path: onboard.py
Line: 183-191

Comment:
**Broken config passes verification**

The generated YAML and JSON checks end with `|| true`, so malformed files still produce a successful exit. `run_verify_gate` treats that exit as a passed check. A newly onboarded config-only repository can therefore report successful verification without any check capable of failing. Preserve parser failures instead of forcing success.

```suggestion
        {"name": "yaml-validate",
         "cmd": "python3 -c \"import yaml, sys; [yaml.safe_load(open(f)) for f in sys.argv[1:]]\" "
                "$(find . -name '*.yaml' -o -name '*.yml' | grep -v node_modules | head -5) 2>&1",
         "cwd": ".", "languages": "yaml"},
    ],
    "json": [
        {"name": "json-validate",
         "cmd": "python3 -c \"import json, sys; [json.load(open(f)) for f in sys.argv[1:]]\" "
                "$(find . -name '*.json' | grep -v node_modules | head -5) 2>&1",
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Comments Outside Diff

These findings could not be posted inline.

•  Scheduled work stops src/tools/gateway_preflight.py:1 ▶

  The move removes root entry points that workflows still call. The required gateway preflight in school-loop.yml still runs gateway_preflight.py, so scheduled execution stops with a missing-file error before reaching the bridge. The CI failure alert also retains the old school_mail import. Update the workflow commands and embedded imports to the new package paths.

•  Build checks stop running src/core/director.py:668 ▶

  Both moved verification callers still pass their own directory as flake_path. These now point to src/core and src/crew, but the only flake.nix is at the checkout root. On a configured Mac runner, run_verify_gate therefore skips every check; strict mode rejects the run. Pass the checkout root from both callers.

•  Root settings stop loading src/crew/crew_dispatch.py:33 ▶

  ENV_FILE still uses the module's parent, so moving director.py and crew_dispatch.py stops them loading the repository-root .env. The crew's _omniroute_api_key fallback has the same problem. If credentials are stored only in the root .env, model calls and spawned crews lose them. Anchor these reads to the checkout root.

•  Curated guidance disappears src/tools/prompt_composer.py:24 ▶

  _CURRICULA_DIR now resolves to src/tools/curricula, while the curated files remain under the root curricula directory. With CURRICULUM_YAML=1, the loader silently returns an empty map and uses static guidance instead. Point the default at the existing root directory.

  _CURRICULA_DIR = Path(__file__).resolve().parent.parent.parent / "curricula"
  

[branben]

Review: 1nzr reorg — mostly solid, but CI is red and several holes need attention

Reviewed against main@9e3ea87, all 294 files, the five commits, and live CI check-runs.

tl;dr

Do not merge yet. The reorg itself is mechanically well-executed (rename detection clean, import sweep verified, no leftover flat imports), but:

1. CI is red on this PR and green on main — pytest (py3.9/3.11/3.12) all abort with a suite-killing timeout. This is a regression introduced here, not pre-existing.
2. 9 Greptile findings confirmed — I verified all of them against the code. The ENTIRE-veto publication gap (#2) and lock-timeout veto-drop (#3) are the most serious.
3. Behavioral changes (new ENTIRE veto, conductor bookbag persistence, new modules) ride inside a PR titled "reorganise ... (reorg-only — no behaviour change)" — the framing is inaccurate and makes review harder.
4. project_verify.yaml coverage regressed — compileall -q src/ no longer covers the six root-level .py files.
5. Runtime data files committed under src/ with machine-specific local paths.

1. CI regression (blocker)

• main HEAD: pytest (py3.9/3.11/3.12) → success (2400 passed, 29 skipped; test_adversarial_review_failure_falls_back takes 30.15s and passes; no timeout marker).
• This PR: all three pytest jobs fail — the run aborts at ~40% with +++ Timeout +++ in tests/test_issue_bridge.py::TestAdversarialReviewStep::test_adversarial_review_failure_falls_back → _apply_lens → _call_with_transport_retry → time.sleep(backoff_s * (attempt + 1)).

Root cause: the PR added @pytest.mark.timeout(10) to a test that inherently takes ≥30s (default LENS_TRANSPORT_RETRIES=2 / LENS_TRANSPORT_BACKOFF_S=5 → 5s+10s sleeps before the fallback raise, plus the model call attempt). On main, without the marker, the global timeout=60 lets it complete at 30.15s. The new 10s budget guarantees the retry sleep exceeds it → pytest-timeout with method=thread aborts the entire run, so none of the ~60% of tests after it ever run.

The PR body's "4 failed, 2474 passed, 34 skipped" doesn't describe CI reality. Either raise the marker to ≥40s, or (better) make the test hermetic — the comment on the patch even admits "This test is not hermetic — if the patch fails... it makes a live HTTP call... and hangs". A hermetic test would patch the transport layer (or inject a fake call_model_fn) so it fails fast deterministically.

(Also: the scan check is red on both main and PR — pre-existing Codex Security Scan, not introduced here.)

2. Greptile findings — all 9 confirmed valid

I verified each against the branch code:

1. _resolve_repo_path fallback returns src/core, not the checkout root — src/core/director.py:1861-1863 Path(__file__).resolve().parent now resolves one level too deep (main's root-level director resolved to checkout root). The reorg fixed sibling path bugs (config_path, REPO_ROOT, CREW_RUNS_FILE, hermes-fm-wrapper) but missed this one. Targeted tests (e.g. test_review_build_gate) monkeypatch _resolve_repo_path, so the fallback isn't exercised. Fix: .parent.parent.parent.
2. CRITICAL ENTIRE veto does not actually block acceptance/publication in the sync bridge path — the veto sets adversarial_review["verdict"]="FAIL", score=0.0, and injects CRITICAL findings, but:
   • acceptance is decided by _review["accepted"] (from the canonical packet/two-judge gate) — untouched by the veto;
   • the combined score is exec*0.5 + review*0.3 + heuristic*0.2, so review=0 still yields a positive score;
   • the PR creation path checks review_evidence (director's review) / the candidate gate — not adversarial_review.
     The bridge's own test asserts status == "success" while also asserting adversarial_review["verdict"] == "FAIL" — i.e. it codifies that a CRITICAL sensor finding still publishes. This contradicts the comment ("matching conductor._persist_acceptance"). Conductor's _persist_acceptance does veto (via the bookbag entire_findings) — the bridge path should mirror that by rejecting the canonical packet / stopping before grading+publication.
3. Lock timeout silently drops the ENTIRE veto — conductor.py:657-664 ignores locked_update_bookbag's None (documented lock-timeout result). If the write times out, _persist_acceptance reads a bag without entire_findings and accepts. The docs say a non-persisted gate is not a gate. Check the return and prevent acceptance when the result wasn't saved.
4. _resolve_lock_entry avoids the online lookup — sources are stored as owner/repo slugs, and the URL expansion only fires when "github.com/" in source. So the ls-remote/gh api repos/{source}/commits branch is dead for collected deps → commit: null even online. Normalize the slug to a clone URL (or drop the guard and use the slug for the API call).
5. Lock entries reused by basename alone — if str(locked.get("name")) == name or locked.get("source","").endswith(source) matches redhat-et/ripwire for a requested acme/ripwire (same basename) and copies the wrong source+commit. Match full normalized source + requested ref.
6. Raw-text scan pre-empts explicit pins — _add uses setdefault, so the regex scan records ref="HEAD" first and the parsed [tool.uv.sources] rev/tag is discarded. Parse structured entries first, or let an explicit pin replace the guessed HEAD.
7. .envit/repos symlinks not verified against the lock commit — _envit_store_path returns the first checkout dir without comparing its commit to the lock, and existing links are skipped on refresh. A stale/different-revision checkout can be linked while the lock claims another pin.
8. yaml-validate / json-validate end with || true — malformed YAML/JSON yields exit 0, so run_verify_gate treats it as a passed check. A config-only repo can report successful verification with no check capable of failing. Drop || true.
9. --require-evidence cannot reject an empty checkout — _collect_evidence always does ev.generated.add("package_manager") even when it only defaulted to npm with no facts read; detected == bool(generated) is therefore always true. Only mark facts actually read.

3. "No runtime behaviour was changed" is inaccurate

The first commit's message and the PR body claim a pure reorg, but the diff adds:

• two new root modules (onboard.py 1718 lines, a2a_bridge.py 272 lines + CLI wiring in issue_bridge.py),
• a new ENTIRE CRITICAL/HIGH veto in issue_bridge.py (not on main),
• a new bookbag persistence block in conductor.py (not on main),
• src/entire_review.py: qodo_replacement → entire_replacement rename and the CLI invocation change (--profile general, --format text dropped),
• review_packet.py::_sensor list/int normalization fix.

Each may be fine on its own, but mixing them into a "reorg-only, no behaviour change" PR makes the reviewer hunt for behavior changes across 294 files. Recommend splitting (or at least renaming) — the reorg alone is already large.

4. project_verify.yaml coverage regression

core-python-compile changed from compileall -q *.py to compileall -q src/. The six root-level modules (a2a_bridge.py, bookbag.py, conductor.py, issue_bridge.py, onboard.py, review_packet.py) — including the two new ones — are no longer compile-checked by the verify gate. A syntax error in issue_bridge.py/onboard.py would now sail through. Either extend to both (src/ *.py at root) or keep a stable root-file list.

5. Committed runtime/test data under src/

• src/sensors/data/activity_log.json (53 lines) — contains local macOS pytest temp paths (/private/var/folders/.../pytest-of-brandonbennett/...) — machine-specific test debris.
• src/tools/data/grading_queue.jsonl — runtime queue data.

The runtime data/ dir is gitignored (data/*), but these live under src/ which isn't. They're runtime/test artifacts that shouldn't be tracked. The .gitignore-visible root data/ is empty of activity_log.json/grading_queue.jsonl in this PR — these are sweep-ups from test runs during the reorg.

Positives

• Rename detection is clean; most renames carry ≤2 non-import changed lines.
• The path arithmetic fixes (config_path, REPO_ROOT, CREW_RUNS_FILE, wrapper path etc.) are correct.
• Import sweep verified: no flat imports of moved modules remain outside .scratch/; all 102 src/ modules + 6 root modules parse cleanly.
• testpaths = tests / norecursedirs in pytest.ini and the stable (non-diff-scoped) ruff scope are good hygiene.
• The noqa-with-reason treatment for the two deliberately-deferred findings (beads tkqn/88jf) is exactly right.

Minimal fix list to unblock

1. Make test_adversarial_review_failure_falls_back hermetic or drop/raise its @pytest.mark.timeout(10) (≥40s) so CI stops aborting.
2. _resolve_repo_path fallback → Path(__file__).resolve().parent.parent.parent.
3. ENTIRE veto: actually reject (accepted=False / stop before grading+publication), and make conductor's lock-timeout path fail closed.
4. project_verify.yaml: compile src/ and the root modules.
5. Remove src/sensors/data/activity_log.json + src/tools/data/grading_queue.jsonl (and gitignore src/**/data/ if runtime files land there).
6. onboard.py #4/#5/#6/#8/#9 — the five mechanical correctness fixes (cheap, well-scoped).

[luna-whymage]

APPROVE — reviewed as the secondary reviewer.

Scope: reorg-only. 93 modules relocated under src/{bridge,core,crew,review,sensors,tools}; the 6 root entry points are retained by design (they are the lint gate's CORE list).

Verified against the tree at bb1b968:

• All 5 path-relative anchors now resolve to the repo root: flake.nix (2155 B), .env (497 B), curricula/, _resolve_repo_path, and context_orchestrator._REPO_ROOT.
• The 6 lint-gate CORE root modules exist at the repo root.
• review-gate passes on this head: 285 changed files, 1402 added lines, no findings.

Note, not a blocker for this PR: .github/workflows/ci.yml:62 names onboard.py and a2a_bridge.py in its ruff CORE list, and neither file exists on this branch or on main. ruff check warns and exits 0 on a missing path, so the gate is green — but the stale list is a silent scope shrink worth a separate cleanup.

Second note, also not a blocker: ci.yml invokes the gate with --head "origin/<sha>", which is not a ref name. The gate read stdout only and discarded the stderr, so it reported 0 changed files and PASSed vacuously — on every PR since #117. Fixing that is a separate change; it does not affect the reorg itself.