Farm inbound-lead mail into the pipeline on the beat (obligations via IMAP → opportunity_sync)

[4444J99]

Context

The inbound-lead farm exists end-to-end but is not connected on the beat, so live recruiter mail (incl. a scheduled interview) sat unfarmed while the daily Opportunity brief reported INBOUND LEADS: 0. Root cause of the dark feed (a variable-name wiring bug, not a dead credential) is fixed in #185 and its companion organvm/limen#1447 (an IMAP --verify probe so a genuinely stale password is caught loudly next time). This issue tracks the remaining integration so the beat actually farms.

Verified findings (2026-07-23)

• With the correct env (GMAIL_USER/GMAIL_APP_PASSWORD), gmail_imap_sweep.py reads all 30 INBOX messages and its MATTER_MAP already buckets recruiter mail as Job Search — but that path is not the obligations source.
• obligations_build.py sources Gmail from audit/inbox_sweep-<account>.json written by the AppleScript inbox_sweep.py, which sees an empty Gmail view (total:0). So no Gmail obligations are built.
• build() (obligations_build.py:220-259) only processes action=="fire" rows and reads r.get("id") for message-ids — but gmail_imap_sweep.classify() rows carry uid (not id) and lack reply_to/snippet, so even if fed, opportunity_sync's message-id dedup (entry_id_for) would collide.
• core/protocols.py inbound-lead-hire matcher (line ~254) keys on recruiter|role at|position at|hiring for|opportunity — it does not catch a calendar-invite interview (e.g. an Ashby/Google-Calendar .ics "Chat with ") nor a bare Needed: <role> subject.

Scope (bounded)

1. Row schema: gmail_imap_sweep.classify() — emit id (= uid), reply_to, and a snippet on each row so obligations/opportunity_sync dedup works.
2. Classify-only receipt mode: a read-only pass that writes audit/inbox_sweep-<account>.json (obligations schema) without archiving, so obligations_build sources Gmail from IMAP, not Apple Mail.
3. Beat wire (organvm/limen metabolize.sh via institutio/governance/sensors.yaml, derive-runner pattern — one registry entry): run the Gmail classify pass → obligations_build → opportunity_sync (application-pipeline). Add an audit/synced_to_pipeline.json receipt so farmed/unfarmed is queryable.
4. Classification (core/protocols.py): classify calendar-invite interviews and Needed:/Seeking: recruiter subjects as inbound-lead-hire (with tests to bound false positives across the corpus).
5. Brief honesty (organvm/limen scripts/opportunity-brief.py): distinguish 0 inbound leads from feed DARK (N accounts returned 0 while inbox non-empty) — a dead feed must read loud, never 0.

Predicate

A done.sh: Gmail feed non-empty when the inbox is non-empty; ≥1 live recruiter thread → inbound-lead-* obligation → applications/<date>/<org>--inbound/entry.yaml; brief reports real count or DARK; idempotent re-run is a no-op.

Related: organvm/application-pipeline#91 (educator/HigherEdJobs discovery inflow — same inflow shape). Companion PRs: #185, organvm/limen#1447.

[4444J99]

Live work is now flowing without waiting on the scheduled beat: the July 28 bounded constituent set was passed through the existing application-pipeline opportunity_sync command on demand, creating six canonical opportunities; the immediate second run returned 0 created / 0 updated / 6 unchanged. Durable receipt: https://github.com/organvm/application-pipeline/pull/93. The remaining scope of this issue is therefore the beat splice and row-contract adaptation (mail classify output into obligation construction and opportunity sync), not the existence or operability of the downstream intake/sync commands.