feat(kimicode): Kimi Code CLI provider, parse ~/.kimi-code wire sessions

[ozymandiashh]

Adds a kimicode provider for Kimi Code CLI (MoonshotAI/kimi-code, the successor of kimi-cli). Requested in #747.

Schema provenance: verified live, not reverse engineered

Unlike the usual new-provider flow, this one was built against a live Kimi Code CLI install (0.27.0, macOS arm64): real sessions were generated locally (success single-turn, multi-tool, multi-turn continuation, retry-only failure), the store was inspected on disk, and the provider was validated against that real store before this PR. The observed schema:

• ~/.kimi-code/sessions/wd_<project>_<hash>/session_<uuid>/ with state.json (title, workDir, ISO timestamps, agents map) and per-agent agents/<id>/wire.jsonl event streams (metadata carries protocol_version: "1.4").
• Per-turn usage arrives as usage.record events: inputOther -> input, output -> output, inputCacheRead -> cacheRead, inputCacheCreation -> cacheWrite. No cost fields exist anywhere, so cost is computed from tokens via calculateCost and flagged estimated (kimi-k3 pricing has been in the fallback since fix(pricing): add kimi-k3 to the pricing fallback #709).
• The alias trap: usage.record.model is the config ALIAS (for example k3), while the real model id (for example kimi-k3) lives on llm.request events. Rows resolve through the alias-to-model map first with the nearest preceding request as fallback, so aliases never leak into reports. This exact case is exercised by a regression test.
• Retry-only failed sessions (repeated llm.request with attempt, no usage.record) parse to zero usage without throwing; malformed JSONL lines are skipped.
• Subagents get their own wire.jsonl under the same session (parsed as separate sources, no double counting); multi-turn continuations append to one wire.jsonl and parse as one session with several turns.
• Tool calls (tool.call loop events) feed the tool breakdown; tool state resets at turn boundaries so a failed turn cannot bleed tools into the next row.

Robustness

Hardened after an adversarial multi-model review pass: alias-first model resolution, turn-boundary tool-state resets, and the dedup key's store invariant (per-agent wire exclusivity) documented in the provider docs.

Validation

• Fixture suite covers: single-turn, multi-tool, multi-turn, retry-only failure, subagent wires, malformed lines, alias-vs-real-model resolution, and tool-bleed regressions. All sanitized, no live store required.
• Real-store check on the machine that generated the sessions: report --provider=kimicode returns 2 usage-bearing sessions, 9 calls, correct token mapping (non-zero cacheRead), and the real model id in the models table; doctor shows the resolved ~/.kimi-code probe root with 4/4 sessions parsed.
• npx tsc --noEmit clean; kimicode + registry + full local suites green.

Scope note

kimi-cli (the predecessor) is being wound down and Kimi Code CLI migrates its config and sessions into ~/.kimi-code on install, so this provider targets the kimi-code store only. KIMI_CODE_HOME overrides the root.

Closes #747