feat: research-informed compaction improvements

[BYK]

Three improvements to the distillation and compression pipeline, informed by insights from the KV cache compression literature (Zweiger et al. 2025, Eyuboglu et al. 2025).

Changes

1. Loss-annotated tool stripping (gradient.ts)

When tool outputs are stripped at higher gradient layers, the replacement now includes metadata — tool name, line count, error presence, and referenced file paths — instead of a static [output omitted] placeholder. This helps the model decide whether to invoke recall for full details.

Before: [output omitted — use recall for details]
After: [output omitted — read: 142 lines, contained errors, paths: src/db.ts, src/gradient.ts — use recall for details]

Inspired by the per-token scalar bias β from "Fast KV Compaction via Attention Matching" (Zweiger et al., 2025) — when tokens are removed, preserving metadata about what was lost helps compensate for information loss.

2. Context-distillation meta-distillation objective (prompt.ts)

The RECURSIVE_SYSTEM reflector prompt now produces a structured working context with sections for Current State, Key Decisions, Technical Changes, and Session Timeline, rather than a flat re-organized event log. This context-distillation objective generalizes better to diverse downstream queries than the prior summarization approach.

Inspired by the Self-Study method from "Cartridges" (Eyuboglu et al., 2025) — the key finding that memorization objectives (produce an event log) don't generalize, while context-distillation objectives (produce context optimized for varied query types) do.

3. Multi-resolution composable distillations (db.ts, distillation.ts, gradient.ts, temporal.ts)

Gen-0 distillations are now archived instead of deleted during meta-distillation:

• Archived entries are excluded from the in-context prefix (no context budget impact)
• Archived entries remain searchable via the recall tool (detail layer for specific queries)
• Archived entries follow the same temporal pruning retention policy (120-day default)

This preserves a multi-resolution view: gen-1 "cartridge" for in-context overview, gen-0 detail for targeted recall.

Inspired by Cartridges composability — independently compressed representations can be concatenated and queried without retraining.

Schema migration

Version 5: adds archived INTEGER NOT NULL DEFAULT 0 column to the distillations table with an index. Backward compatible — existing rows default to archived = 0.

Test results

All 187 tests pass (only change: schema version assertion updated from 4 to 5).