[gardener/classifiers] raise DIFF_CAP and filter noise — 20KB is 6% of Haiku's window

[serenakeyitan]

Problem

Both classifiers (anthropic.ts:40, claude-cli.ts:36) cap diff input at 20,000 bytes with no content awareness. Two issues compound:

1. 20KB is far too small. The default model is claude-haiku-4-5 (200K context window). Current total prompt is ~55KB (~13K tokens) — only 6% of available context. The cap truncates real PRs at 300-400 lines of code. A typical feature PR is 800-1500 lines; a refactor or migration is multiples of that.

2. Lockfile noise eats the budget. A PR that regenerates pnpm-lock.yaml blows through 20KB of lock diff before a single line of real code reaches the classifier. tree sync already solved this (sync.ts:665 formatPrDiffForPrompt) by filtering noise first, then truncating.

Proposal

1. Raise caps

Constant	Current	New	Share of 200K window
DIFF_CAP (both classifiers)	20,000	200,000	~25%
DIGEST_BUDGET_BYTES (tree-digest.ts:25)	30,000	100,000	~12%
Combined with system + metadata	~55KB	~310KB	~38% of Haiku's 200K

Leaves ~60% of window for prompt-caching overhead, tokenizer variance on mixed Chinese/code content, and Anthropic's soft ~180-190K prompt-length threshold. Works for Sonnet/Opus upgrades without re-tuning.

2. Filter noise before truncating

Drop from both classifiers' diff input:

• *.lock, package-lock.json, pnpm-lock.yaml, yarn.lock, Cargo.lock, poetry.lock, Gemfile.lock
• *.min.js, *.min.css, *.map
• dist/, build/, node_modules/, coverage/, __pycache__/, out/

Regex list already exists in sync.ts:626-630. Extract to shared engine/classifiers/diff-filter.ts so sync and both classifiers share one source of truth.

3. Single default — no flag

One set of numbers for all scenarios (daemon sweeps, CI, manual invocations). Keeping it simple; we can add an override later only if a real workload demands it.

Cost implication

Haiku 4.5 input pricing is cheap enough that raising per-classify cost from ~$0.004 → ~$0.02 is a non-issue for the verdict quality gain. Prompt caching (#315) keeps repeated classifies on the same tree cheap.

Acceptance

• DIFF_CAP raised to 200,000 in both anthropic.ts and claude-cli.ts
• DIGEST_BUDGET_BYTES raised to 100,000 in tree-digest.ts
• Noise filter extracted to shared module and applied in both classifiers before truncation
• Existing gardener-*-classifier.test.ts still pass, updated fixtures if needed
• One new test: lockfile-heavy PR produces non-empty real-code diff in prompt