examples: document deferred follow-ups (metrics, fixtures, pushshift)

docs/plans/2026-05-28-reddit-examples-followups.md

@@ -0,0 +1,40 @@
+# Reddit examples — follow-ups (deferred from initial plan)
+
+Items deliberately deferred from
+`docs/plans/2026-05-28-reddit-examples-plan.md` so the initial cut can
+ship without disproportionate scaffolding.
+
+## Prometheus counters (Task 11a)
+
+The design doc names four counters:
+
+- `fire_examples_scraped_total{sub}`
+- `fire_examples_extracted_total{sub,confidence_bucket}`
+- `fire_examples_llm_fallback_total`
+- `fire_examples_extract_failed_total{reason}`
+
+Current state: log-based observability only. The ingest CLI emits one
+"ingest done: inserted=X skipped=Y sub_runs_succ=Z/T" line per run.
+
+Reason for deferral: the ingest runs once a week via CronJob. Wiring
+`prometheus_client` Counters into a non-FastAPI Job requires either a
+pushgateway hop or a metrics-emitting HTTP endpoint exposed for the
+Job's lifetime — both are disproportionate plumbing for a weekly
+batch. Revisit if we ever need drift signal (e.g., extraction quality
+collapsing silently) or higher cadence.
+
+## Fixture expansion
+
+The design said 20 hand-curated fixtures; Task 12 ships with 5 (one
+per fi_status / currency combination that exercises a distinct code
+path). Add the remaining 15 once the live ingest has run — pull
+candidates from the real DB by hand and lock them in as regression
+anchors.
+
+## Pushshift / pullpush.io for older posts
+
+PRAW's `top-of-all-time` caps at 1000 posts per subreddit. For the 12
+target subs that's ~12k posts before deduplication. If a Q3-2027
+review of `fire_planner.fire_example` shows we're consistently missing
+milestone posts older than the cap, layer pullpush.io as a secondary
+source feeding the same pipeline.