fix(poller): distinct URLs are distinct leads + source diagnostics (1.7.7)

[CryptoJones]

What

Ports a code review from the sibling applytrack repo. Two of the five findings apply here; the other three are already handled by the .NET/Postgres architecture (verified, no change made).

Finding 1 — dedup dropped distinct openings (the core fix)

score_and_stage deduped scraped listings by a normalized company + role slug that stripped trailing parentheticals, so per-city variants of one role — each a distinct application URL — collapsed to a single lead and the rest were silently recorded as seen and dropped. (In the sibling repo this staged 1 of five LawnStarter city postings and dropped four; the same logic lived here.)

• _norm_slug now keeps the full role title (parentheticals included).
• Seen.has makes the posting URL authoritative — a distinct URL is a distinct lead; the slug is only the fallback key for URL-less listings.
• score_and_stage records seen keys per outcome (staged / filtered / link-dead) instead of up front, and no longer marks a listing seen when staging raises, so a genuine write failure is retried next run.
• PollRepo.add_lead suffixes -N on a slug-name collision (two distinct postings whose titles normalize to the same name) instead of dropping the second.

Verified end-to-end: 5 distinct city URLs → 5 staged; re-poll → 0 (no re-pings).

Finding 4 — source failures were swallowed

A dead/renamed job source looked like an empty market. poll._gather and worker._gather_by_source now log each failing source at WARNING (and tolerate any exception, not just HTTP/parse); the worker also logs the per-tenant setup/poll failures it had been swallowing.

Not applicable here (verified, no change)

• Finding 2 (non-atomic file writes): the poller is Postgres-backed and writes no state files.
• Finding 3 (optimistic-concurrency clobbers): the draft endpoint writes a separate cover_letters table, not the versioned application row; the blacklist batch (PassOpenLeadsAsync) is a single atomic UPDATE of status/version/updated_at guarded by status IN ('lead','ready'), so it preserves unrelated fields and won't flip an advanced row — there's no read-modify-write race to guard, and adding one would reduce safety.
• Finding 5 (stale docs): README/SPRINTS already describe poll as discovery-only.

Tests

ruff + mypy --strict + bandit clean; pytest green — dedup tests updated for URL-authoritative behavior, plus new distinct-URL, unique-name, url-less-slug, and source-failure-logging cases. dotnet build of the API clean. Version 1.7.6 → 1.7.7 (patch) in pyproject.toml + ApplyTrack.Api.csproj.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W84TwzhCdhmoLoxXYV2NuH