ops(registry): backfill stale member_profiles.agents types + crawler reclassify-on-disagreement

[EmmaLouise2018]

Refs #3538.

PR #3498 added resolveAgentTypes() server-side, but it only runs on writes (POST/PUT to /api/me/member-profile). Rows saved before #3498 never get re-evaluated. The crawler's type-update path at crawler.ts:580 only wrote back when the stored type was missing — once any non-unknown value was set, the row was frozen, so Bidcliq and Swivel (registered as 'buying' while being sales agents) cannot self-correct.

This is the cleanup for Problem 1 in #3538.

Crawler type-update policy (server/src/crawler.ts)

Old: write back only when no stored type and inferred is non-unknown.

New:

• Promote when stored is missing OR stored is 'unknown' AND inferred is non-unknown. Same intent as before, broadened to cover the 'unknown' case that was previously frozen.
• Log on disagreement (stored non-unknown ≠ inferred non-unknown). Do NOT auto-flip — single probes can be wrong; auto-flipping would corrupt good rows on a transient bad probe. Operator runs the backfill explicitly to reconcile.

Backfill script (server/scripts/backfill-member-agent-types.ts)

Walks every member_profiles row, calls resolveAgentTypes() on its agents[], writes back any agent whose stored type disagrees with the snapshot's inferred type. Idempotent. Has a --dry-run mode.

Operator runbook

Owner: member-tools / registry oncall. Runs from a workstation with DATABASE_URL set to the target environment (staging first, then prod). The script is read/write — same machine that runs migrations.

Expected diff size today:

• Bidcliq, Swivel are the known cases: 'buying' → 'sales'. Both are member-registered, both have agent_capabilities_snapshot rows that infer sales. Two flips minimum.
• Possible additional flips: any other member-registered agent saved before fix(registry): server-side agent-type resolution at registration #3498 whose snapshot disagrees with its stored type. Order of magnitude: < 10 across the current member fleet.
• If the dry-run reports > 25 flips, stop and investigate before running for real. Something else is going on (probe drift, mass mis-classification, classification-inference change). Operator escalates rather than absorbing the diff.

Where output goes:

• Stdout from the script — capture to a file for the run record (e.g., 2026-04-29-staging-backfill.log).
• Paste the captured stdout into a comment on this PR after the staging run, and into the prod-deployment ticket / runbook entry after the prod run.
• Future audit ("when did Bidcliq flip?") relies on this paper trail until the audit-log table from ops(registry): backfill writes type-reclassification audit log #3550 lands.

Procedure:

# 1. Dry-run on staging
DATABASE_URL=<staging> npx tsx server/scripts/backfill-member-agent-types.ts --dry-run | tee 2026-04-29-staging-dryrun.log

# 2. Eyeball the diff. Expected ≈ Bidcliq + Swivel + 0–8 long-tail.
#    If >25 disagreements or any unexpected agent_url surfaces, STOP.

# 3. Real run on staging
DATABASE_URL=<staging> npx tsx server/scripts/backfill-member-agent-types.ts | tee 2026-04-29-staging-real.log

# 4. Spot-check the public registry — Bidcliq + Swivel render as sales now.

# 5. Repeat 1–4 against prod with prod DATABASE_URL.

Pre-run checklist:

• Dry-run output captured.
• Dry-run diff matches expected set ± documented surprises (no unexpected agent_url, no > 25 flips, every flip explainable).
• Surprises documented in PR comment / ticket before proceeding.
• Real run only after staging run validated and prod backup window aligns with normal practice.

Export

resolveAgentTypes is now exported from server/src/routes/member-profiles.ts so the script can reuse it. Pushing the abstraction up rather than duplicating it across the script + the write path.

Test plan

• New: server/tests/unit/crawler-type-update-policy.test.ts — pins the promote/disagreement matrix. 5/5 pass.
• npx tsc --noEmit -p server/tsconfig.json — clean.
• Pre-commit hooks green.
• Operator: run --dry-run on staging per the runbook above.

Stack ordering

Recommended merge order from #3538: 3540 → 3542 → 3543 → 3541. Backfill ships last so it's the most-deliberate step — runs against the wire-corrected codebase from #3540 and the docs-explained surface from #3542 / #3543.

Out of scope

• ops(registry): backfill writes type-reclassification audit log #3550 — type-reclassification audit log (the structured replacement for the stdout-capture protocol above).
• feat(registry): one-shot re-probe of all type='unknown' agents with extended timeout (Problem 2 quick-win) #3551 — one-shot re-probe for unknown agents (Problem 2 quick-win).
• decision(registry): how should discovered agents surface member-publisher endorsement (#3538 Problem 6) #3547 — Problem 6 decision (discovered-agent member-publisher endorsement).
• Public registering-an-agent.mdx page — docs(registry): registering-an-agent — four crawl paths + AAO-membership value story #3543.

[bokelley]

Issue #3550 proposes a type_reclassification_log table (or reuse of an existing audit-log pattern) — same surface as this PR (backfill-member-agent-types.ts + crawler disagreement-log path). You've already noted it as out of scope in the PR body, so this is just the cross-link: consider folding before merge or confirm as follow-up post-merge.

---

Generated by Claude Code