feat(profiles): approve source origins for Microsoft, Mistral AI, xAI, and Cohere

[abdeslam-menacere]

Approves source origins for four long-tail creators — Microsoft, Mistral AI, xAI, Cohere — so that a later data refresh can cite them. Closes #437.

This populates nothing. The /tree/ page will still render "No non-featured creators in the reviewed catalog." after this merges, correctly. It unblocks the refresh that comes next.

What changed

Five files, additions only, all under a new tools/updater/profiles/origins/ directory:

file	what it approves
microsoft.json	microsoft.ai, www.microsoft.com (bounded to /en-us/research/blog/)
mistral-ai.json	mistral.ai, docs.mistral.ai
xai.json	x.ai, docs.x.ai
cohere.json	cohere.com, docs.cohere.com
README.md	why the directory exists, and what was deferred

Trusted origin union: 16 → 24. The two huggingface.co entries add no new origin; they bound link-following.

Why a subdirectory

The placement is the design, and it is load-bearing. Three consumers read tools/updater/profiles/, and they enumerate it differently:

consumer	enumeration	sees origins/?
gate-source-approval.mjs → catalogAnchor()	git ls-tree -r (recursive)	yes — this is what makes the change work
gate-evidence.mjs:165	readdirSync, skips non-files	no
profiles.py:364 → _reviewed_profile_paths()	iterdir(), skips non-files	no

So the origins are visible to the trust anchor and invisible to the reviewed-creator sets. A top-level microsoft.json would instead break test_profiles.py's EXPECTED_CREATORS and promote Microsoft to the 2-of-3 pilot review bar, which #437 places explicitly out of scope. The precedent already exists: profiles/generic/long-tail.json lives in a subdirectory for the same reason.

QA proved the skip is load-bearing rather than incidental by hoisting origins/microsoft.json to the top level: discovery count went 4 → 5, and load_profile_library then refused with ProfileError - microsoft.json: missing required field 'creator'. The worst-case layout mistake fails closed rather than silently promoting a creator.

Gate verdicts

Both gates ran against b57a729, read issue text and git diff only, and never saw the developer's summary or session. Neither reviewed its own work.

gate	verdict	record
review	PASS	comment 5445591330
qa	PASS	comment 5445754879, addendum

drydock is not installed in this environment, so no machine gate receipt exists and none has been hand-written. The issue comments are the authoritative record.

⚠️ The required check passes without executing anything

Correction, made after this PR was opened. An earlier revision of this description claimed CI would run nothing at all for this change. That was wrong, and the error was mine: I enumerated web-ci.yml and skills-ci.yml, reasoned from those two, and never listed the full workflow set. updater-tests.yml exists, triggers on tools/updater/**, and is running the pytest suite on this PR against Python 3.11 and 3.13. Leaving that claim standing would have understated the coverage a reviewer actually has.

The accurate position is narrower, and still worth reading before trusting the checks:

workflow	runs here?	why
web-ci	skips	predicate ^(web/|\.github/workflows/web-ci\.yml$) — and it is the only required check on main
skills-ci	skips	predicate ^(\.github/skills/|\.github/scripts/|\.github/workflows/skills-ci\.yml$|web/src/data/)
updater-tests	runs	paths: tools/updater/** — the Python suite, on 3.11 and 3.13

So the Python half is covered by CI. Two real gaps remain:

1. The required check is satisfied vacuously. web-ci will go green having executed nothing for this change, so branch protection is not evidence here.
2. The node gate suite is not run at all. gates.test.mjs is where gate-source-approval.mjs — the script this change actually depends on — is exercised, and skills-ci skips it. skills-ci.yml documents this against itself at lines 110–117: "a pull request confined to that directory skips gates the change would have failed. A green skip records that the gates did not run, never that they would have passed. Do not read one as coverage."

The node gate results and the web result below were therefore run by hand, because nothing in CI runs them for this change. The pytest figures are reproduced by updater-tests on this PR.

Real output

Run against main + this branch merged, re-run after main advanced (see below):

node --test gates.test.mjs        -> tests 149, pass 149, fail 0, exit 0
check-skill-doc-test-counts.mjs   -> OK, 8 markdown file(s), exit 0
tools/updater: python -m pytest   -> 1095 passed, 2 skipped, 3 deselected, exit 0
web: npm.cmd run validate         -> 24 files, 603/603 passed, astro 0 errors, exit 0

The 3 deselected are test_live_publication.py (2) and test_network_provider.py (1), excluded by pre-existing pyproject.toml:65 config (-m 'not network and not live') because they perform real egress. Not hidden coverage.

npm run validate was run against the earlier merge base. git diff origin/main HEAD -- web/ is empty — this branch changes no file under web/ — so main's own CI result carries unchanged.

Mutation probes (QA) — every one tripped

probe	expected	observed
delete origins/ from the merge	union shrinks	24 → 16, 8 accepts → 16 failures, exit 1
8 accept bundles at the pre-change anchor	all refused	8/8 exit 1
14 unapproved origins post-change	all refused	14/14 exit 1 — incl. Azure, learn.microsoft.com, subdomain squat, scheme swap, userinfo spoof x.ai@evil…
promote any of the four to pilot	must not	still long-tail; openai control exits 0, so the probe is not vacuous

main moved twice under this branch

Both gate verdicts bind to b57a729, which is unchanged, so they still hold. But the merge target moved from a052811 → 0556ec7 → 29acbda, and two of those commits touch machinery this change depends on. Rather than carry a stale result forward, everything was re-measured against the current tip:

• fix(gates): count profile catalogues read, not profiles listed #443 rewrote 73 lines of gate-source-approval.mjs — the exact script this change relies on. Verified it does not affect trust: its own docstring states "urls is unchanged by that bookkeeping. Which origins this anchor approves is decided exactly as before." It splits profileFiles from profileCatalogues for reporting.
• feat(ci): verify the web build on every pull request and push to main #442 rewrote 137 lines of web-ci.yml. Verified the scope predicate is byte-identical, so the skip above still applies.
• The gates baseline moved 147 → 149 because fix(gates): count profile catalogues read, not profiles listed #443 added two tests. The figure above is the current one.
• Union re-measured at main 29acbda: still 16 → 24, same 8 origins.
• git merge-tree --write-tree against 29acbda: exit 0, clean. No file in this branch overlaps anything main changed.

Deferred rather than guessed

• DeepSeek — api-docs.deepseek.com unreachable at verification time, a connection failure rather than a 4xx, so absence of evidence was not recorded as evidence of absence.
• Alibaba / Qwen — the creator entity is genuinely unresolved.
• Azure and learn.microsoft.com — deliberately withheld. Microsoft is simultaneously a creator (Phi, MAI), a serving platform (Azure), and a product vendor (Copilot); collapsing those would violate the entity-boundary rule. github.com stays trusted as code hosting only, not as Microsoft's announcement channel.
• Two conflicts recorded rather than smoothed: x.ai now self-describes as "SpaceXAI LLC" (verified against the live <title> and og:title), and huggingface.co/CohereLabs is isVerified: false, so it was declined.

Known gaps, disclosed

1. No committed test validates the content of origins/*.json. Set membership is guarded — test_profiles.py:35 asserts EXPECTED_CREATORS against the real directory and fails on a leak (proven: 6 errors under mutation, clean on revert). But the new urls, trust_notes and allowed_paths are unchecked by any suite. A future edit adding azure.microsoft.com would be caught only by human review.
2. allowed_paths is advisory at this gate. Trust attaches at scheme://host; nothing in tools/updater/src/ parses these documents. Disclosed in the README rather than implied to be enforced.
3. catalogAnchor swallows parse errors with continue. Pre-existing and fails closed, but newly load-bearing now that a second directory feeds it.
4. cohere-docs and mistral-docs record a url outside their own allowed_paths. Investigated and benign — meta.json and google-deepmind.json already do the same; it is the established "root entry point, link-following bounded to checked subsections" convention.

Merge

This is outside ADR 0003's qualifying class — gate-scope.mjs reports all five files outOfClass, "passed": false, exit 1, so it cannot auto-merge and must not. Auto-merge has deliberately not been enabled.

The merge is a human's. The required check (web-ci) will pass without executing anything for this change, so please weigh the hand-run node-gate output above and the updater-tests result rather than the green tick alone.