feat(v3.0): wire API-based summarizer from config (supersedes #26)

[RichardHightower]

Summary

Replaces #26 with the same feature (wire ApiSummarizer into the daemon from config + env vars, replacing the hardcoded MockSummarizer TODO) plus the review feedback from #26 applied. All four task pr-precheck gates pass: fmt, clippy --all-targets --all-features -D warnings, full workspace test (~825 tests), and doc with RUSTDOCFLAGS=-D warnings.

What's the same as #26

File	Change
crates/memory-daemon/src/commands.rs	build_summarizer() reads provider/model from config, resolves API key from env var or config, constructs ApiSummarizer or falls back to MockSummarizer with warning.
crates/memory-types/src/config.rs	api_key_env: Option<String> added to SummarizerSettings with serde(default) for backward compat.

What's improved over #26 (review-driven)

Decision logic split (functional core / imperative shell)

Five of #26's seven new tests ended with let _ = summarizer; — they only proved build_summarizer() didn't panic. They couldn't detect a regression that always returned Mock or always returned Api. This PR extracts a pure decision function:

pub(crate) fn pick_summarizer_kind(
    settings: &SummarizerSettings,
    key_available: bool,
) -> SummarizerKind  // Mock | OpenAi | Anthropic

Tests now assert on the exact SummarizerKind returned for every provider × key-availability combination.

API hygiene

• Drop dead 2-tuple from resolve_api_key. Was fn resolve_api_key(...) -> (Option<String>, &'static str) but the only caller did let (api_key, _) = .... The discarded second element was also misleadingly the default env var name even when api_key_env override was active, so any future caller using it would log the wrong var on a cache miss. Now fn resolve_api_key(...) -> Option<String>.
• Downgrade pub → pub(crate) on build_summarizer, env_var_for_provider, pick_summarizer_kind. memory-daemon has both lib.rs and a [[bin]] target, so pub was leaking test helpers as part of the library's public API.

Helpers + duplication

• Extract is_anthropic / is_openai / default_env_var_for_provider using eq_ignore_ascii_case (was: 3 sites of .to_lowercase() == "anthropic" with per-call allocations).
• warn! on unknown providers before fail-open to OpenAI (was: silent fall-through — typos like "openi" or different vendors like "google" would silently route to api.openai.com).

Tests: 7 weak → 16 strong

Group	Count	Notes
env_var_for_provider mapping	4	Kept as-is
pick_summarizer_kind decision	6 (new)	Assert exact SummarizerKind for openai / anthropic / mixed-case / unknown / no-key paths
resolve_api_key resolution	5 (new)	Includes precedence regression guard (api_key wins over api_key_env even when env var is set with a different value)
build_summarizer smoke	1	Verifies wiring + Arc::strong_count == 1

The precedence test is the one #26's description claimed but didn't actually have.

Why a new PR instead of force-pushing #26

#26 also added a "v3.0 — Real Summarization / Phase 51: API Summarizer Wiring" section to ROADMAP.md and STATE.md. Local main (not yet on origin/main) has v3.0 Competitive Parity & Benchmarks with Phase 51 = Retrieval Orchestrator, Phase 52 = Simple CLI API, Phase 53 = Benchmark Suite — so #26's ROADMAP edits would collide on phase numbering once local main is pushed. This PR drops the planning edits in a separate revert commit; the summarizer work will be slotted into the existing v3.0 plan in a follow-up (likely a new Phase 54 "Real Summarization", or folded into Phase 51's LLM reranker work).

CI expectations

All checks should mirror #26 (which was fully green): Format, Clippy, Test (ubuntu+macos), Build (ubuntu+macos), Documentation, E2E Tests, and the full E2E CLI matrix across claude-code / gemini / opencode / copilot / codex on both ubuntu-24.04 and macos-latest.

Closes #26