v0.6.0 — Phase 3: Measurement

Phase 3 makes forge's core thesis — the harness makes small models capable — measurable and regressable.

Highlights

• forge bench — task-eval harness reporting harness-lift (same model bare vs full) with per-lever ablation across 8 switchable levers (schema, workspace, plan-pin, loop-detect, read-gate, alias-repair, escalation, compaction). Default = all levers on = byte-for-byte unchanged.
• Structured step trace — one schema-versioned meta record per session + one step record per loop iteration; the formerly-invisible malformed/loop/compaction events now log durably. New forge trace <sid|last>.
• Flight recorder + forge replay — records raw model output per step (malformed raws included) into a RecordingBackend cassette; re-drives a real Agent.send with no model, turning any recorded session into a zero-inference regression fixture.
• Agent-loop invariant battery — direct tests for escalation, loop detection, compaction, plan pinning, and inbox absorption.

Suite: 215 tests green, stdlib-only, zero runtime deps.