feat(kimi): estimate session cost from aggregate token usage

[Junt184]

Kimi sessions previously showed no cost and never appeared in the usage
views. Kimi wire logs record token counts only as session-level
aggregates (the StatusUpdate path), so individual messages carry no
per-message token_usage and no model identifier. The cost engine
priced 0 tokens per message and produced $0 for the whole session.

This change emits a single session-level ParsedUsageEvent when a Kimi
session exposes only aggregate output tokens, and routes parser-emitted
usage events through processKimi (which previously dropped them).

The cost it produces is explicitly an estimate, not exact billing:

• Prices come from the LiteLLM catalog (a community-maintained price
  list), not from Kimi/Moonshot directly.
• The wire logs do not record which model served each turn, and Kimi
  sells several models concurrently at different rates (e.g.
  kimi-k2.7-code, kimi-k2.6, kimi-k2.5), so the per-turn model is
  unknown. The event defaults to moonshot/kimi-k2.6, a recent model
  with a cleanly matchable catalog entry. k2.7 is currently priced only
  under a cloudflare-namespaced key that the provider-conflict rule
  rejects for a moonshot/ model, and it carries the same 0.95/4.0 rate
  as k2.6, so k2.6 is the more robust proxy at an identical price. Bump
  the constant once the catalog gains a cleanly matchable k2.7 entry.
• Only output tokens are available from these aggregates, so input and
  cache tokens are not priced. The figure is a lower bound that tracks
  output volume rather than a full invoice.

Sessions whose messages already carry per-message token_usage (the
native step.end protocol) are still priced message-by-message and are
skipped here to avoid double counting.

Reviewers may want to look at defaultKimiModel in
internal/parser/kimi.go for the estimation rationale and the
emission/skip logic at the end of ParseKimiSession.

[roborev-ci]

roborev: Combined Review (c7ae53b)

Medium issues found; no High or Critical issues reported.

Medium

• internal/parser/kimi.go:808 - dataVersion was not bumped after adding persisted Kimi usage events. processKimi skips unchanged files whose stored data version is current, so existing synced Kimi sessions will not be reparsed or backfilled and will still show no estimated cost until the source file changes or a forced resync occurs.
  Fix: Bump dataVersion in internal/db/db.go and update the corresponding version guard test/comment so existing archives resync.

• internal/parser/kimi.go:530 - currentModel now defaults to defaultKimiModel, and step.end assigns that before checking whether the event has its own model. Native per-message usage rows with an explicit event.model but no prior config.update can be stored and priced as the proxy model instead of the actual model.
  Fix: Prefer event.Get("model").Str in the step.end path before falling back to currentModel or defaultKimiModel, and add coverage for that case.

---

Panel: ci_default_security | Synthesis: codex, 8s | Members: codex_default (codex/default, done, 4m4s), codex_security (codex/security, done, 26s) | Total: 4m38s

[roborev-ci]

roborev: Combined Review (03e3bd9)

No issues found.

---

Panel: ci_default_security | Synthesis: codex | Members: codex_default (codex/default, done, 4m57s), codex_security (codex/security, done, 14s) | Total: 5m11s