🧾 parse: expand usage.iterations into per-model attempt rows

[martinciu]

🧭 Summary

Closes #456.

Billed usage.iterations entries served by a different model than the turn's — refused fallback attempts and cross-model advisor calls — were silently folded into the serving model's totals. This PR expands informative iterations at parse time into per-model attempt rows (message_id keyed <id>:it:<idx>), so every aggregation surface (models panel, projects, buckets, Recost) attributes tokens and cost to the model that actually consumed them.

🔧 What changed

• 🧩 pkg/parse — new expandIterations + toMessages: own-model entries fold into the parent (iterations are authoritative when informative — also repairs the rare near-zeroed/under-reported outer usage), each non-zero foreign entry becomes an attempt row inheriting the full envelope. Fail-open paths (undecodable blob, oversized probe, missing message.id) keep today's behavior byte-for-byte.
• 🔁 Both parse entry points wired — the reader path (ParseWithErrors) and the incremental tail path (ParseFromOffsetWithErrors) that the indexer/watcher actually use. The tail path was initially missed and caught by the real-data differential below; a regression test now pins it.
• 🗃️ pkg/cache — SchemaVersion "11"→"12" with zero schema-text changes: the bump only forces the wipe-and-rebuild that backfills attempt rows from historic JSONL on first launch. Quota history (usage_samples/usage_limits) survives, as pinned by the new migration test.
• 🧪 Pinning tests — attempt rows flow through InsertMessages idempotently, redistribute in ModelAggregates, and re-price against their own model's rate in Recost — all with no SQL changes (design claim held on first run).

📊 Real-data differential (live JSONL, ts < 2026-08-06T07:00Z)

Baseline binary (pre-change) vs this branch, both cold-indexing the same tree:

Model	Δ tokens (in+out)	Δ cost	Why
claude-opus-4-7	+1,433,310	+$26.23	17 attempt rows + parent folds
claude-opus-5	+1,373,747	+$8.23	4 attempt rows + parent folds
claude-fable-5	+441	+$0.83	2 attempt rows (refused fallbacks)
claude-opus-4-8	+379	+$0.54	parent folds (served 4 informative turns)
claude-sonnet-4-6	0	+$0.07	under-reported cache_write_5m repaired on 17 informative parents
claude-sonnet-5	0	+$0.05	same repair, 5 parents
every other model	0	$0.00	byte-identical

Totals: 124,970 → 124,993 rows (+23 attempt rows, matching the 40 informative turns on this machine), +2,807,877 tokens, +$35.95 — previously misattributed or invisible.

✅ Testing

• 🧪 go test ./... — all packages pass
• 🧹 make lint — 0 issues
• 🔬 Real-binary smoke: index --rebuild differential above; cost-only movers individually verified as informative-turn parents with repaired cache columns

📣 Surface changes (for release notes batching)

• Schema v12: first launch after upgrade wipes and re-indexes the message cache from JSONL (quota history preserved). Attempt rows appear as extra messages rows keyed <id>:it:<idx>; per-model totals shift accordingly.

🤖 Generated with Claude Code