Add burn claude-md: CLAUDE.md hot-path cost attribution

[willwashburn]

Summary

• New burn claude-md command: per-section cost attribution for the project's CLAUDE.md, with cross-session aggregation and a per-session cost breakdown
• New burn claude-md advise subcommand: emits unified-diff TRIM hunks for the most expensive sections (read-only — burn never mutates CLAUDE.md)
• Detects CLAUDE.md at the project root and at `.claude/CLAUDE.md`

Closes #10.

Output shape

```
CLAUDE.md at /Users/will/Projects/foo/CLAUDE.md (412 lines, ~3.1k tokens)

Cost per session: avg $0.18, p95 $0.34
Cost over 7d: $4.12 across 23 sessions

Sections ranked by cost:
lines heading tokens cost/session %file
12- 47 ## Architecture 1.8k $0.11 59.0%
48- 89 ## Testing conventions 890 $0.05 28.7%
90- 148 ## Tone and personality 410 $0.02 13.2%
```

Attribution math

Each turn's CLAUDE.md cost is computed directly: `claude_md_tokens × cacheReadPrice / 1M` per turn where `cacheRead >= claude_md_tokens`. This simplification of the issue's share-based formula falls out cleanly because CLAUDE.md size is constant across the session — the share-vs-direct difference vanishes. The `cacheRead < claude_md_tokens` check skips:

• The first turn (CLAUDE.md is in `cacheCreate`, not `cacheRead`)
• Turns where the file has been compacted out of cache

Section attribution is proportional: `section_cost = total_cost × (section_tokens / total_tokens)`.

Section parsing

• Groups by H2 if any exist (matches the standard CLAUDE.md shape: H1 doc title + H2 sections)
• Falls back to H1 if no H2s
• Treats a heading-less file as a single "preamble" section
• Skips fenced code blocks so example markdown inside them isn't parsed as headings
• "Preamble" content (above the first grouping-level heading) is its own section

Test plan

• `pnpm run build` — clean
• `pnpm run test` — 193/193 passing (11 new tests)
• Synthetic-attribution test (5 turns × known CLAUDE.md size) within ±10% of hand-computed truth
• Section ratio test: section costs are proportional to token shares
• "Skip when cache too small" test
• Empty CLAUDE.md returns zero cost
• Section parser tests: preamble, H2 grouping, H1 fallback, code-fence skip
• `findClaudeMdFiles` finds root + nested
• Advise diff test: hunk header line range matches the actual section

Notes / follow-ups

• Per-session-time-travel (CLAUDE.md as it existed at session time, via git log) is out of scope for MVP — current implementation always uses HEAD's CLAUDE.md. Documented in the README via the `since` flag note.
• Nested `src/**/CLAUDE.md` discovery is not yet implemented; only root and `.claude/CLAUDE.md` are scanned. Easy to add when sessions in the wild start using deep CLAUDE.md trees.
• Rollup across worktrees via `projectKey` is already wired (`Query.project` matches both `project` and `projectKey`)

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 5 additional findings in Devin Review.

[devin-ai-integration]

🟡 Fence closing check matches lines with trailing content, incorrectly exiting code blocks

The closing fence check at line 176 uses trimmed.startsWith(fenceMarker.charAt(0).repeat(fenceMarker.length)), which matches any line that begins with enough backticks/tildes — even if the line has trailing content (e.g. ````python). Per the CommonMark spec, a closing fence must contain only fence characters and optional whitespace. When a line like ````python appears inside a 3-backtick code block, this code incorrectly closes the fence, causing subsequent headings that are still inside the code block to be detected as real sections. Meanwhile, the actual closing ``` line is misinterpreted as opening a new fence, swallowing headings after the real code block.

Reproduction

Input:

inside block

## should-be-inside
```
## should-be-outside
```

Expected headings: only `## should-be-outside`.
Actual headings: only `## should-be-inside` (wrong one).

</details>

```suggestion
      } else if (new RegExp(`^${fenceMarker.charAt(0) === '`' ? '`' : '~'}{${fenceMarker.length},}\\s*$`).test(trimmed)) {
```

<!-- devin-review-badge-begin -->
<a href="https://app.devin.ai/review/agentworkforce/burn/pull/51" target="_blank">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1">
    <img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open in Devin Review">
  </picture>
</a>
<!-- devin-review-badge-end -->

---
*Was this helpful? React with 👍 or 👎 to provide feedback.*