feat(operations): quota-independent context planning with deterministic shrink #24

Motivation

planConversationContext returns requires-portable-handoff whenever the latest compaction is native-only (Codex encrypted compacted record) and the target is another provider, regardless of size, and requires-compaction whenever the effective history exceeds the target budget. Both outcomes demand a live turn on the source provider. When the source has exhausted its subscription window, that turn cannot happen, so exactly the sessions that most need switching cannot switch. A census of the last 300 local Codex rollouts shows 146 with a compacted record.

The raw pre-compaction records remain on disk in plaintext for both providers; only Codex's summary is encrypted (server-held key, no client-side decryption exists in codex-rs). Anthropic's context editing and OpenAI's own Claude-session importer both shrink transcripts deterministically (clearing or narrating tool results) without a model call.

Intended behavior

• planConversationContext(conversation, target, budget, { allowSourceTurns: false }) never returns a plan that needs a source turn. New outcomes: raw-history (native-only compaction entries dropped, raw entries fit) and shrunk (deterministic shrink applied; carries a report), plus a hard unfittable error only when the last complete user turn alone exceeds the budget.
• New isolated module operations/shrink.ts: strip native-only compactions → clear tool results oldest-first (bounded placeholder, keep recent turns intact, edit inputs preserved) → drop oldest complete user turns with a synthetic compaction entry that indexes the dropped prompts. Pure, reported, no provider names inside.
• compactionAvailability returns rejected for a Claude carrier whose text starts with a rate-limit prefix (You've hit your, You've used, You're now using extra usage, You're close to, You're out of extra usage). Hosts must never accept such a record as a portable summary.
• Claude assistant records with isApiErrorMessage: true decode as opaque entries instead of assistant messages, so a limit error is never projected into the target as model output.
• Existing outcomes and the default allowSourceTurns: true behavior are unchanged.

Acceptance criteria

• Unit tests run against redacted real fixtures extracted with the existing observed-fixture tooling: a Codex rollout with one and with several compacted records, a Claude transcript with a compact boundary, a Claude transcript containing a rate_limit API error record, and an oversized transcript for the shrink ladder.
• Structural projection tests prove Claude and Codex native projectors accept cleared tool results and the synthetic drop marker.
• The live resume probe gains a shrunk-projection case (opt-in, real headless resume with a read-only marker prompt).
• No change to any provider's native persistence rules in docs/design/provider-switching.md other than the new planner outcomes.