docs: clarify audit doc-Lows + record won't-do disposition (PR #94)

[lesnik512]

What

PR 5 of the tail program — documentation only, no code changes. Adds the clarifications the pass-3 audit flagged and records the final disposition of every remaining finding.

CLAUDE.md (invariant nuances)

• F2-09 — within-batch dispatch order is unspecified (ORDER BY governs selection, not the per-row dispatch sequence).
• F1-01 — the true lease bound is fetch_batch_size + max_workers, not fetch_batch_size.
• F2-04/F2-05 — activate_at / publish_batch NOTIFY timing is worker-clock-relative (vs activate_in's server-side make_interval).
• F1-07 — NOTIFYs emitted during a fetch-loop reconnect window are lost (latency, not correctness).
• F2-07 — deliveries_count counts claims, not handler runs; size max_deliveries with margin.
• F2-01 — max_total_delay_seconds is a lower bound (allows ~one more interval).
• F4-09 — test-broker session leniency on publish/cancel_timer/fetch_unprocessed.

docs/operations/troubleshooting.md

• F3-03 — "Recommended alerts" (outbox row count, lease_lost rate, dlq_written vs nacked_terminal divergence) on the unbounded-growth section.
• F5-03 residual — new entry: a chained-OutboxResponse publish failure surfaces as a handler-failure nack with no distinct signal; how to tell them apart (+ index row).

planning/audits

Appended a Tail disposition ledger: resolved (#85–#94), documented (#94), won't-do with rationale (F4-03, F6-06, F1-02, F2-02-residual, F5-03-residual, F5-04, F3-04), deferred (#95: F7-07, F7-09, F1-08).

Verification

• just lint clean (ruff + ty)
• mkdocs build --strict passes (no broken anchors/links)
• non-Postgres tests green (no code touched)

After this, only the deferred test-hardening items (#95) remain from the entire pass-3 audit.

🤖 Generated with Claude Code