feat: configurable compaction model

[santhreal]

Adds modelSystem.compaction, a setting that selects the model used to summarize the conversation during context compaction (both automatic compaction and /compact).

• Defaults to the main conversation model (inherit); set { "backend": "...", "model": "..." } to compact with a different model (e.g. a cheaper or longer-context one).
• Prompt-budget sizing still tracks the main model — that is the context being fit; only the summary call uses the compaction model.
• If the configured backend isn't ready (e.g. missing API key), compaction inherits the main model and prints a warning in the compaction notice (warn-and-inherit).
• Wired through all three compaction paths: /compact, session auto-compact, and the in-loop guard.

Also bundles a small clippy question_mark cleanup in hooks/state.rs.

Tests: 5 new (model_system round-trip + context_guard classify/notice), full suite green, clippy --all-targets -- -D warnings clean, fmt --check clean.

[morganlinton]

Thanks for this — clean design (budget stays on the main model; only the summary call switches), and the follow-ups for /compact warnings, /route discoverability, and failure surfacing are appreciated.

A few small nits before merge:

1. Docs wording (CHANGELOG.md / README.md) — Prose says set modelSystem.compaction to a "provider:model" spec, but the actual config shape is a ModelRef object ({ "backend": "...", "model": "..." }). The JSON example is correct; please reword the surrounding text so it matches (e.g. “a ModelRef / { backend, model } object”).

2. CompactResult.warning doc comment — Still describes only the unavailable → inherit-main-model case. After the failure-surfacing commit it also covers dedicated-model error/empty → deterministic trim. Please broaden the comment to match both paths.

3. Missing Unavailable unit test — classify_compaction_model covers Inherit and ready Use; please add a case for a configured backend that isn’t ready (e.g. openrouter without an API key) asserting CompactionModel::Unavailable.

Happy to discuss separately whether a failed dedicated compaction model should retry with the main model before falling back to deterministic trim.

[santhreal]

Thanks Morgan — all three addressed in 16fdc36, plus the retry idea:

1. Docs wording — CHANGELOG and README now describe modelSystem.compaction as a ModelRef object ({ backend, model }, same shape as planner/selector) instead of a "provider:model" spec. JSON examples were already in that form.
2. CompactResult.warning doc — broadened to cover both paths: unavailable backend → inherit main model, and a dedicated model that errors/returns-empty → retry → deterministic trim.
3. Unavailable test — added compaction_marks_not_ready_backend_unavailable, which forces OPENROUTER_API_KEY empty (then restores it) and asserts CompactionModel::Unavailable with the key-required reason.

On the retry discussion — I went ahead and implemented it since it's clearly the better behavior: a dedicated compaction model that errors or returns empty now retries with the main conversation model before falling back to the deterministic trim. The notice reports the outcome accordingly ("… failed; summarized with <main> instead", or "… and main model both failed; trimmed deterministically"). Budget sizing still tracks the main model throughout. Happy to gate that behind a flag if you'd rather it be opt-in.

All green: 8 compaction tests (incl. the new one), clippy --all-targets -- -D warnings, and fmt --check.

[morganlinton]

Looks good — review items addressed, and the main-model retry before deterministic trim is the right default. Approving.