… + drift placeholders; warn on budget exhaustion

Addresses #343 items 1 & 2 (budget growth + noise filtering). Leaves
items 3–5 (leaf inclusion, soft_links, task-aware model selection,
onboarding.md doc) for follow-up PRs since they involve more design.

## What

Three focused changes to `tree-digest.ts`:

1. **DIGEST_BUDGET_BYTES: 100KB → 500KB.** Today's Claude 4.5/4.6/4.7
   models give us a 200K-token (~800KB) context window; the old
   budget was severely over-conservative and silently truncating
   nodes from the classifier's view on larger trees. 500KB covers
   ~2700 NODE.md entries with room for PR body + diff + response.
2. **SKIP_DIRS gains `.gardener-tree-cache`.** Gardener's own
   per-sweep snapshot of the tree — stale duplicates of real NODE.md
   files — should never enter the digest. On the paperclip-tree
   repo this alone cut the digest node count roughly in half.
3. **Drop `drift/` auto-generated placeholder nodes by summary
   regex.** These are scaffolding nodes gardener writes to give
   proposal PRs a valid parent chain (summary: "Auto-generated
   intermediate node for sync proposals"). They carry zero decision
   signal — safe to filter before they eat budget.

Plus one observability change:

4. **`collectTreeDigestDetailed()` reports `skippedAsNoise`,
   `truncatedCount`, and `budgetExhausted`.** The two classifier
   factories (`claude-cli`, `anthropic-api`) now call the detailed
   variant and emit `emitDigestDiagnostics()` output via an
   injectable `write` sink (defaults to `process.stderr`). Before:
   budget exhaustion was silent → verdict could cite wrong nodes
   as "closest match" with no trace. Now:

       gardener: tree digest filtered 5 drift placeholder node(s) (auto-generated by prior sync)
       gardener: tree digest budget exhausted — 17 node(s) dropped. Verdict may miss relevant tree context. Consider pruning the tree or raising DIGEST_BUDGET_BYTES.

## E2E results on paperclipai/paperclip + paperclip-tree

Before/after on the real tree (v0.3.2 main → this branch):

|                          | before | after  | delta            |
| ------------------------ | ------ | ------ | ---------------- |
| digest nodes kept        | 138    | 64     | −74 (−54%)       |
| digest bytes             | 24,794 | 11,188 | −55%             |
| `.gardener-tree-cache`   | in     | out    | eliminated       |
| `drift/` placeholders    | in     | out    | 5 filtered       |
| budget headroom          | 100KB  | 500KB  | 5×               |

Ran `gardener comment --pr 4367 --repo paperclipai/paperclip` end-to-end
against the built CLI:

- Diagnostic line fired (`filtered 5 drift placeholder node(s)`).
- Classifier completed against the cleaner digest.
- Verdict sharpened from NEEDS_REVIEW (v0.3.2) → NEW_TERRITORY,
  which is the more precise call: PR #4367 adds a queue-sweep
  governance behavior no existing tree node covers. The cleaner
  digest + removal of stale cache duplicates let the classifier
  confidently place the PR in "new territory" instead of hedging.

## Tests

- New `tests/gardener/tree-digest.test.ts` (10 tests):
  - `.gardener-tree-cache` is skipped.
  - Drift placeholder summary is filtered; same-word mentions in real
    summaries pass through (anchored regex).
  - Detailed-result accounting stays self-consistent.
  - `emitDigestDiagnostics` emits the right lines (noise only,
    budget only, both, or silent on a healthy digest).
  - `formatDigest` bullet format unchanged.
- All existing gardener tests pass (380/380 in `tests/gardener/`,
  including the 11 claude-cli classifier tests and 19 anthropic
  classifier tests).
- Typecheck clean.

## Scope

Deliberately does not ship items 3–5 of #343:

- Leaf-file inclusion when PR diff paths overlap a node domain.
- `soft_links` transitive inclusion.
- Task-aware default model (haiku for comment, sonnet-4-6 for
  sync/draft-node).
- `GARDENER_CLASSIFIER_MODEL` documented in onboarding.md.

Those involve more design (heuristic tuning, per-command plumbing,
doc edits) and are better reviewed separately.