Kiro: IDE v2 session store support + credit-based pricing across all formats

[Enclavet]

Summary

Kiro's newly released 1.0 IDE (observed on 1.0.138; the 0.12.x builds still write the v1 globalStorage format) moves sessions to ~/.kiro/sessions/<workspace-hash>/sess_<id>/ — a self-contained session.json + messages.jsonl event log — instead of globalStorage. This PR adds a v2 parser and discovery, and — validated against a real upgraded machine — moves Kiro cost accounting from token estimation to the real metered credits Kiro records, across every format that carries them (CLI, v1 IDE executions, v2 IDE).

Token estimation wasn't just imprecise, it inverted model rankings. On a real machine, one week of Kiro usage:

Model	Token-estimated	Metered credits	Error
Sonnet 4.6	$339.82	$161.52	2.1× overstated
Opus 4.6	$11.15	$130.57	11.7× understated
Fable 5	$7.20	$118.15	16× understated

What's included

New: IDE v2 store (sess_*/messages.jsonl)

• Parses the event-sourced turn stream (user / turn_start / assistant / tool_call / tool_result / usage_summary / turn_end) with per-execution dedup keys (kiro-v2:<session>:<execId>) and defensive flushes for out-of-order events and in-flight sessions.
• Prices turns from usage_summary.promptTurnSummaries[].usage credits at the public overage rate (USD_PER_KIRO_CREDIT = $0.04; individual plan is $20/mo for 1,000 credits) — same never-understate approach as the Codebuff provider.
• Counts tool_result content as input context (matching the CLI parser's ToolResults treatment) and takes the real modelId from session.json, so v2 sessions aren't mislabeled kiro-auto.
• v1 → v2 is a clean cutover (no migration/dual-write observed); dedup namespaces are disjoint, so nothing is counted twice.

Fixed: credit pricing now actually reaches reports

• metering_usage values are credits, not dollars. The existing CLI parser summed them as USD, overstating CLI cost 25× at the real rate.
• Kiro joins the costUSD pass-through allowlist in providerCallToCachedCall (alongside mistral-vibe, devin, hermes, vercel-gateway, antigravity). Without this, the parser's credit-based cost was discarded at the cache boundary and re-derived from estimated tokens — silently undoing all credit pricing. Tradeoff (shared with all allowlisted providers, documented): price-override / model-alias don't affect Kiro dollar amounts.
• v1 IDE executions are credit-priced too. usageSummary[].usage (the v1 predecessor of v2's promptTurnSummaries) was already being iterated for usedTools but its credit values were ignored. Since v2 only ships in brand-new builds, v1 is the format nearly all Kiro IDE users are on today. Verified per-turn semantics on real data (values fluctuate; not cumulative counters).
• Uniform fallback contract. All three credit parsers gate on summed credits > 0 and fall back to token-estimated pricing with costIsEstimated: true. This also fixes two CLI edge cases: no-metering turns were priced at a frozen $0, and empty metering_usage arrays (observed on 75 of 10,105 real turn metadatas — meta written before metering lands) passed the old truthy-presence check and froze $0 marked as real.

Fixed: v2 reasoning tokens double-counted

Reasoning was folded into outputTokens and reported in reasoningTokens, but aggregation sums the two fields. Now disjoint (matching codex/gemini); reasoning is still billed as output in the pricing fallback.

Cache invalidation

Both persistence layers froze pre-fix dollar values and never self-invalidate (historical session files don't change):

• session cache: CACHE_VERSION 4 → 5
• daily rollup: DAILY_CACHE_VERSION 10 → 11 with MIN_SUPPORTED_VERSION raised (same pattern as the v10 cursor accounting bump)

One-time full re-parse; verified live that the corrected numbers land.

Testing

• 20+ new tests: v2 parsing/discovery/dedup/routing, credit vs fallback pricing per format, tool_result turn scoping, empty-metering edge, reasoning disjointness, and a mixed-format coexistence suite (legacy .chat + v1 executions + workspace-sessions + CLI + v2 on one machine) asserting exact aggregate counts and no double counting.
• The cost pass-through test is mutation-verified: it runs end-to-end through parseAllSessions and fails with the token re-price ($0.000048 vs $0.10) if the allowlist entry is removed.
• Validated against a real machine: 9,995 CLI turns, 221 v1 executions, and 24 v2 turns now carry metered cost ($1,793 total); call counts unchanged before/after (no calls gained, lost, or double-counted).
• Full suite green; tsc clean; tsup bundle builds.

Docs

docs/providers/kiro.md updated: v2 store layout, credit pricing sources per format, dedup namespaces, the parse-time cost-freezing contract (and its price-override implication), and the disjoint-store v1→v2 verification. A comment at fingerprintFile documents the companion-file blind spot (CLI credits / v2 modelId live next to, not in, the fingerprinted transcript).