fix(compaction): count compact_boundary markers instead of inferring from context-% drops

[CorticalCode]

Why

The Compaction Counter (#282) infers compactions from drops in used_percentage between renders. The heuristic is fragile: it samples a noisy, debounced signal and guesses about discrete events. Known failure: transient used_percentage: 0 frames create false counts — #370 (thanks @KorenKrita) patches that symptom with a sub-1% guard, but the problem class is inherent to inference: missed events between renders, window-size changes, and persisted prev-state that can go stale.

What

Count the explicit {"type":"system","subtype":"compact_boundary"} records Claude Code writes to the session transcript at every compaction. Exact, stateless, retroactive. Records with isSidechain: true (a subagent compacting its own context) are excluded, matching the main-chain convention already used in jsonl-metrics.ts. Display is unchanged (↻ N). The heuristic, its zod schema, threshold logic, and the per-session cache under ~/.cache/ccstatusline/compaction/ are deleted — net −319 lines. Marker format verified stable across transcripts from v2.1.92 → v2.1.170, for both auto and manual triggers (v2.1.170 adds extra compactMetadata fields, but type/subtype/isSidechain are unchanged).

Behavior notes

• Stateless and file-scoped: concurrent Claude Code instances count independently, and the count accumulates across --resume/--continue (markers live in the session's transcript file).
• Gated on transcript_path, like every other transcript metric.
• Adds one transcript read per render when the widget is enabled — the same per-metric pattern as token/duration/speed metrics. (A shared single-read across metrics would be a nice follow-up.)
• Old per-session cache files under ~/.cache/ccstatusline/compaction/ become inert; they're left in place, consistent with how the existing per-session skills cache is handled.
• The 0%-flash false-count class can't occur by construction.

Testing

• 10 unit tests (TDD): exact counting, 0%-flash immunity, sidechain exclusion, malformed lines, missing/unreadable transcripts.
• Full suite: 1369 pass / lint clean / bundled build verified.
• End-to-end: piped renders against fixtures and real-world transcripts, cross-checked against grep -c marker counts — old code shows a false ↻ 1 on a flash frame; this branch shows ↻ 0 (no markers) and exactly ↻ 2 (two markers + flashes + a sidechain marker).