release: v0.6.0 — RAG singleton, retry/backoff, parallel polish, on-disk polish cache

[silversurfer562]

Summary

Four perf and resilience improvements on the polish/RAG path, plus a new cache CLI subcommand for the on-disk cache that gets introduced.

• rag_hook: process-level RagPipeline singleton (thread-safe via threading.Lock + double-checked locking). Corpus loads once per process instead of once per template kind — measurable win on --all-kinds runs.
• doc_gen/_anthropic: retry with 1s / 2s / 4s exponential backoff for 429, 529, and APIConnectionError. Non-retryable SDK errors raise immediately. Credential redaction and __cause__ stripping preserved.
• generator: three-phase render → polish → write. Polish is now ThreadPoolExecutor-parallel (max 4 workers) so wall-clock time drops to roughly the slowest single LLM call instead of the sum.
• polish: on-disk cache at ~/.attune/polish_cache/ (overridable via env). Key includes content + source_summary + template_type + system_prompt + augmented_context + model, so any input or model change invalidates entries. _cache_get bumps mtime on hit so heat is observed reliably even on noatime mounts. Lazy mtime-based prune (default TTL 30d, env-tunable, 0 disables) piggybacked on _cache_put. Manual nuke via attune-author cache clear.

Why

Re-runs of attune-author regenerate --all-kinds were paying full LLM cost even on unchanged source. The cache, plus parallel polish, plus prompt caching downstream in attune-rag#3, bring the bill and the wall-clock down by an order of magnitude on the warm path. The retry/backoff and singleton are robustness wins independent of cost.

New CLI surface

attune-author cache clear    # delete every cached polish entry

New env vars

Var	Default	Purpose
ATTUNE_AUTHOR_POLISH_CACHE	~/.attune/polish_cache	Cache directory override
ATTUNE_AUTHOR_POLISH_CACHE_TTL_SECONDS	2592000 (30d)	Mtime TTL; 0 disables prune

Version

0.5.1 → 0.6.0 (minor — additive but new behavior + new subcommand).

Dependencies

• Allows but doesn't require attune-rag#3 (>=0.1.0,<0.2 pin satisfies 0.1.10).
• Allows but doesn't require attune-help#4 (>=0.10.0 pin satisfies 0.10.0).

Test plan

• tests/test_polish_cache.py — new, 12 tests: hit/miss, mtime bump on hit, model in key, prune by mtime, TTL=0 disables, invalid TTL falls back to default, clear_cache, end-to-end polish_template cache hit skips LLM
• tests/test_anthropic_retry.py — new, 9 tests: 429/529/APIConnectionError retries, exponential schedule, gives up after _MAX_RETRIES, non-retryable raises immediately, credential redaction, __cause__ stripped
• tests/conftest.py — autouse fixture resets RagPipeline singleton between tests
• Full suite: 518 passed, 37 skipped (was 497, +21 new tests)
• Smoke-tested attune-author cache clear against the live cache (deleted 16 entries)
• Smoke-tested TTL prune via _cache_put against a 1-second TTL — hot entries survive, expired entries die

🤖 Generated with Claude Code