feat(runtime): batteries-included compression policies (truncating / extractive / anchored)

[roryford]

Summary

MK already ships the CompressionPolicy / PreTurnCompressionPolicy seams (wired into ConversationRuntime via TurnCompressionCoordinator), but no concrete strategy — the protocol docs show a hand-wave compress() stub and every consumer had to write their own. This PR adds three batteries-included strategies + one policy wrapper.

Ported and rewritten from Fireside's StoryCompression subsystem, adapted from its internal Message/Result tuples to MK's [ChatMessage] seam (generate is a parameter, not stored → clean Sendable structs, no @MainActor/@unchecked).

Strategies (a ladder)

• TruncatingCompressionStrategy — zero-inference sliding window (keep system + newest-by-budget tail, drop oldest). The canonical baseline. (new — not in Fireside)
• ExtractiveCompressionStrategy — zero-inference scored selection (recency / length / keyword-density), verbatim tail, greedy within budget. Adds an optional headBudgetFraction knob to pin establishing context (middle-out / lost-in-the-middle).
• AnchoredCompressionStrategy — summarize old messages via generate, prepend a .memory("summary") record + keep a verbatim recency tail. Keeps Fireside's summarizerInputWindow input-sizing decoupling, chunk-and-fold for over-window input, summary-floor logic, CancellationError early-return, and fall-back-to-extractive on failure/empty/missing-generate.

Policy wrapper

DefaultCompressionPolicy is a Sendable struct conforming to both CompressionPolicy (post-turn) and PreTurnCompressionPolicy (pre-turn). The trigger asymmetry is handled internally — the post-turn path uses contextUtilization; the pre-turn path receives only messageCount/lastPromptTokens, so the policy stores contextSize as config to compute utilization.

runtime.compressionPolicy        = .extractive(threshold: 0.75)
runtime.preTurnCompressionPolicy = .anchored(threshold: 0.85, contextSize: 8192)

Tests

Tests/ManifoldRuntimeTests/DefaultCompressionPolicyTests.swift — ported from Fireside's StoryCompressionTests + P0 extractive-edge-case / anchored-fallback / summarizer-starvation suites, rewritten against [ChatMessage], plus new truncating-strategy, headBudgetFraction, and pre-turn-vs-post-turn trigger tests.

Out of scope (deliberately)

• Recursive / hierarchical (summary-of-summaries) summarization
• Embedding / semantic-relevance selection (belongs with RAG)
• KV-cache eviction (token/attention layer, not message history)
• The graph reconciler postCompress hook (left as the default no-op for consumers to fill)

Notes for review

• Verify the isPinned resolution against ChatMessage (Fireside's Message had isPinned; check whether MK exposes a pin/kind equivalent or whether v1 treats none-pinned as acceptable with a TODO).
• Draft pending CI + a review-and-fix pass.

🤖 Generated with Claude Code

[roryford]

Review-pass note (from the adversarial review + fix pass, commit 477830ba):

• Deliberate divergence from the Fireside original: ExtractiveCompressionStrategy drops Fireside's explicit candidateBudget = budget - tailBudget cap and fills candidates to the full history budget. This is intentional — strictly better budget utilization, still bounded, never overflows context. Not a port bug.
• isPinned: no per-message pin equivalent exists on the [ChatMessage] seam (pins live on ChatSession.pinnedMessageIDs, invisible here). The port preserves .system + .memory records verbatim as 'load-bearing' and leaves a TODO to honor a future pinned-IDs channel.
• Known coverage gap (not a bug): the anchored chunk-and-fold over-window path (summarizerInputWindow set + input exceeds it) has no direct test.
• Local gate: scripts/test.sh --profile local → 4932 tests, 0 failures, 15 skipped, clean first run.