feat(crossposting): source-quality ranker + diversity cap (Phase 2)

[ohld]

Summary

Phase 2 iteration on the channel crossposting ranker — published spec (specs/completed/crossposting-source-quality.md) plus full implementation, tests, and review.

Behavior changes:

• Per-source quality multiplier (0.5×–2.0×) using AVG(forwards × √(views/100)) over mature 30-day crossposting history per meme_source_id
• Self-calibrating median normalization (PERCENTILE_CONT(0.5)) — set-and-forget, no hardcoded thresholds
• Diversity cap: ≤1 post per source per 24h on each channel, filtered to confirmed posts (telegram_message_id IS NOT NULL)
• In-bot share boost: (1 + min(invited_count, 10) × 0.1) — caps at ×2.0 for memes with 10+ user shares
• Image-only filter in source-quality CTE (excludes April video-era confound)
• Bot-engagement quality floor (nlikes >= 5) preserved per channel premise: "best of bot, not newest of source"
• score_version=2 tag on new posts → enables clean v1/v2 comparison

Production correctness fixes (from codex review iters):

• src/flows/rewards/uploaded_memes.py: moved log_meme_sent to after bot.send_media_group so reward posts now record real telegram_message_id and properly participate in the diversity cap
• Failed pre-send rows no longer poison the cap (telegram_message_id IS NOT NULL filter)
• log_meme_sent keeps ON CONFLICT DO NOTHING semantics — preserves the original mature crossposting sample when reward flows repost an already-crossposted meme

Graceful degradation:

• Ranker returns None when filters exhaust the pool → flow logs warning and skips slot (no Sentry alert)
• Cold-start safety: if no source has 5+ mature image posts, source-quality multiplier falls through to neutral 1.0

Verification

• ruff check src/ tests/ — clean
• ruff format src/ tests/ — clean
• pytest tests/test_crossposting_meme.py — 13/13 pass (4 new ranker tests + 9 existing)
• All SQL CTEs verified on prod data:
  • RU: 7 qualifying sources, median signal 33.35, 5 sources blocked by cap, top-5 candidates valid
  • EN: 4 qualifying sources, median 1.67, 4 blocked, top-5 valid
• Prod health pre-merge: 0 app errors / 0 Prefect failures / 0 Sentry crossposting issues

Success measurement

After this lands, baseline before next changes:

• RU: 16.1 avg forwards/post (n=73 mature, last 14d)
• EN: 2.7 avg forwards/post (n=61)

Re-read at +14d post-ship. Pass = v2 ≥ 1.15× v1 forwards on at least one channel + no >10% drop in views. Query in spec.

Review trail

• /plan-ceo-review — selective expansion mode, scope/strategy
• /plan-eng-review — architecture, tests, performance
• Codex outside-voice (8 findings, 5 accepted, 1 rejected on user product principle)
• ralphex — 4 implementation tasks + 9 codex iterations + 2 claude review rounds
• Final codex iter-9: NO ISSUES FOUND

Test plan

• All 13 unit + integration tests pass (host-mode pytest)
• All ranker SQL CTEs return non-empty results on prod data
• No regressions in pre-merge prod health
• Post-merge: trigger Post to TG Channel RU flow once, verify a meme posts cleanly
• Post-merge: trigger Post to TG Channel EN flow once, verify a meme posts cleanly
• Post-merge: monitor Sentry + prod logs for 30min after first successful posts

Deferred (TODOS)

• VK measurement layer
• Channel-share deep-link signal (sc_%) materialized into meme_stats
• Re-test videos as soft penalty post-baseline (after 2026-05-12)
• Drop dead caption IS NULL multiplier

[ohld]

STAFF ENGINEER REVIEW: APPROVED — pure docs + gitignore PR. Spec doc + .worktrees/ ignore. /review + /codex review both clean. /cso skipped (no sensitive surfaces touched).

One non-blocking note: the spec has two duplicated 'final SQL' blocks with different nlikes thresholds (>= 5 vs >= 2) — confusion risk for the implementation PR (ralphex), not a merge blocker.

CI: lint pass, test pending — --auto will gate the squash-merge until test passes.