feat(claude): account for /advisor escalations under the advisor model

[iamtoruk]

What

Claude Code's new advisor tool (/advisor) escalates hard decisions mid-turn to a stronger model. Those tokens are recorded inside message.usage.iterations[] as advisor_message entries carrying their own model, and are excluded from the top-level message.usage totals (which cover only the main model).

CodeBurn read only the top-level totals, so for anyone using /advisor:

• the advisor's tokens were silently dropped (undercounted spend), and
• no cost was ever attributed to the advisor model (wrong model mix).

This is the same gap ccusage closed in their advisor support.

Change

• Parse each advisor_message iteration into its own call, attributed to the advisor's model and priced accordingly. The main call keeps exactly the top-level totals, so there is no double counting.
• Preserve advisor iterations through two paths that would otherwise drop them:
  • entry compaction (compactEntry), which trims usage for the session cache;
  • the large-line (>32KB) byte-scanner parsers, which never extracted iterations before. The fix JSON-parses only the small usage-object slice, not the whole line.
• Fold advisor cost into the live guard budget (src/guard/usage.ts), which also reads these sessions.
• Bump the claude parser cache version and GUARD_CACHE_VERSION so existing caches re-parse and pick up advisor spend.

Advisor calls use a stable dedup key (<msgId>:advisor:<ordinal>) keyed on advisor ordinal, so the report path (compacted, advisor-only iterations) and the guard path (raw iterations) agree.

Verification

• New tests in tests/parser-advisor-usage.test.ts: main-call attribution unchanged, separate advisor call priced under the advisor model, no double counting, survives compaction, and survives the >32KB large-line path. Mutation-verified each fails on pre-fix code.
• Adversarial review (two rounds) executed streaming/dedup, cache round-trip, guard-vs-report parity, UTF-8 in the large-line buffer path, and small/large parity; all clean.
• Full suite: 3094 tests pass. tsc --noEmit clean.

Scope note

fallback_message iterations (written when a turn retries on another model) are a related, pre-existing under-count in the same family. They were never counted before this change either, so they are left as a documented follow-up rather than expanding this PR.