fix(cache): CacheManager protocol, async-only guard, clear key errors, InMemoryCache max_size LRU (v26.06.15)

[ancongui]

Summary

An edge-case audit while validating the implement-cache-strategy skill (which validated clean — cache hits provably skip the source via an observable counter, and evict invalidates) surfaced real framework bugs in untested corners.

Fixed

1. CacheManager didn't satisfy its declared CacheAdapter protocol. It implemented only get/put/evict/clear, so isinstance(mgr, CacheAdapter) was False and using one as a @cacheable backend raised AttributeError: 'CacheManager' object has no attribute 'exists' on the first null-cached hit. Added exists / put_if_absent / evict_by_prefix / start / stop (mirrored to primary + fallback).
2. Cache decorators silently broke on sync functions. They await an async backend, so a sync target failed with a cryptic await TypeError at call time. They now raise a clear TypeError at decoration time.
3. Bad key templates raised a bare KeyError at call time; now a clear ValueError naming the unknown parameter.

Added

• InMemoryCache(max_size=...) with LRU eviction, wired from pyfly.cache.max-size. The adapter was unbounded (the max_size stat was always None — advertised but unimplemented); default stays unbounded.

Documented (by-design, not bugs)

• Cache keys are namespaced by the backend instance + key template — reuse a template across methods only for the same logical entry (this is what lets @cache_evict invalidate a @cacheable entry; qualname-namespacing would break that pattern).
• The Redis JSON round-trip is lossy (a cached Pydantic model returns as a dict on a Redis hit, unlike the in-memory adapter).

Tests / Gates

tests/cache/test_cache_hardening.py (7): manager protocol + decorator-backend null-caching, sync-rejection, bad-key error, LRU eviction + unbounded default. mypy --strict (607) ✓ · ruff + ruff format ✓ · full suite 3681 passed, 1 skipped.

Bumps v26.06.14 → v26.06.15, CHANGELOG, uv.lock synced.