fix(claudecode): include cache tokens and per-model window in [ctx: ~N%]

[AaronZ345]

Two bugs broke the [ctx: ~N%] indicator in the claudecode path:

1. Cache tokens missing — handleResult read only usage.input_tokens (per-turn delta). In long sessions almost all input is cache_read_input_tokens, so the delta stays <100, sdkPlausible := event.InputTokens >= 100 is false, and the suffix never appears.

2. Hardcoded 200k window — contextIndicator divided by const modelContextWindow = 200_000. Users on claude-opus-4-7[1m] saw 5x inflated percentages.

agent/codex already does the equivalent (reads total_token_usage + model_context_window from rollout events). This PR brings claudecode to parity.

Changes

• core/message.go: add Event.ContextWindow int (additive).
• agent/claudecode/session.go: new modelContextWindow(model) maps [1m]/-1m → 1M, default 200k. handleResult sums input_tokens + cache_read_input_tokens + cache_creation_input_tokens and stamps the window on the event.
• core/engine.go: remove the constant, rename to defaultContextWindow (fallback only). contextIndicator(inputTokens, window int) takes window as a parameter.

Tests

• TestHandleResultAggregatesCacheTokensAnd1MWindow — verifies three-field sum + 1M window for [1m] model.
• TestModelContextWindow — table test across Opus/Sonnet/Haiku variants, empty string, uppercase.
• Existing TestHandleResultParsesUsage / TestHandleResultNoUsage unchanged and green.
• go build ./... clean.

Related: #547

[chenhg5]

Review Summary

Fixes two real bugs in the [ctx: ~N%] indicator for Claude Code sessions. Well-tested and consistent with the existing codex implementation.

---

✅ Good:

• Correctly aggregates input_tokens + cache_read_input_tokens + cache_creation_input_tokens — matches how Anthropic bills context
• modelContextWindow() cleanly detects [1m]/-1m suffix (case-insensitive)
• ContextWindow field on Event is a clean extension — no breaking changes
• contextIndicator() gracefully falls back to defaultContextWindow when window <= 0
• Two focused test additions: cache aggregation test and 7-case model window table test
• Consistent with agent/codex which already does the equivalent

---

LGTM. Both bugs verified, tests comprehensive, CI green.

[AaronZ345]

Closing — #774 subsumes both fixes from this PR:

• The per-model context window ([1m] → 1M) is reproduced verbatim as claudeContextWindow in feat: structured per-turn reply footer with platform-specific small-text styling #774.
• The cache-token aggregation in handleResult is replaced by a more accurate per-sub-call snapshot in handleAssistant, matching the codex precedent (LastTokenUsage over TotalTokenUsage) and avoiding the long-agentic-turn ctx % clamp at 100% (see @Cigarrr's transcript measurements in feat: structured per-turn reply footer with platform-specific small-text styling #774).

Thanks @Cigarrr for the catch.

[AaronZ345]

Closing — superseded by #774. The per-model context window fix ([1m] → 1M) is preserved in #774's claudeContextWindow(), and the cache-token summing into ctx % was the wrong fix (see #774 review thread for the result.usage aggregation issue and the codex LastTokenUsage precedent). Thanks @Cigarrr for the careful catch.