Measured: 2026-07-03 · CoalHearth v1.0.0 · workers = Sonnet 5, K=3 per arm · fidelity
arm run through the real lib/state-snapshot.js.
TL;DR: on a 10-file mid-refactor interruption, warm resume (CH recovery block) and cold restart both finished correctly (4/4 remaining refs found, 0 redone) and the token delta was <1%—at this scale a strong model reconstructs state from the tree, so CH's token saving is a large-session effect, not a small-session win. CH's irreducible value is state fidelity: its journal preserves the goal, the modified-file set, and—the piece a cold restart can NEVER reconstruct—the in-flight sub-agents (CH 1, cold 0).
| Arm | Found all 4 remaining? | Redid a done file? | Missed the doc-ref? | Mean tokens (K=3) |
|---|---|---|---|---|
| COLD (re-scan all 10) | ✅ 3/3 | none | none | 39,584 |
| WARM (CH block + 4 files) | ✅ 3/3 | none | none | 39,303 |
Token delta: −0.7% (281 tokens). Both arms 100% correct on this fixture. The delta is swamped by the fixed per-agent baseline (~39k) at a 10-file scale—the re-scan CH avoids is only ~6 small files here. On a large repo the cold re-scan-to-reconstruct is what balloons; this toy cannot show that, and we do not extrapolate a number.
Honest read: at small interruption scale, cold restart is not materially worse on tokens OR correctness with a strong model—CH's completion-time win is a scale effect. Its value at any scale is the fidelity arm below.
buildStateSnapshot() on the mid-refactor state, with one in-flight sub-agent recorded:
| State item | CH recovers | Cold restart recovers |
|---|---|---|
| Goal / task | ✅ yes | guess from the tree |
| Modified-file set | ✅ 6 (journaled exactly) | git-status guess (mixes pre-existing) |
| Remaining work | ✅ the checklist | infer from the tree (lossy at scale) |
| In-flight sub-agents | ✅ 1 | ❌ 0—no tree signal exists |
The irreducible finding: a sub-agent that was spawned and then lost to the interruption leaves zero trace in the working tree—a cold restart cannot know it ever ran. Only CH's PostToolUse journal records it (Incident E). CH's honest scope holds: it records that the worker existed + its residue path, so a human/agent can find and reconcile it—it does not recover the worker's lost compute. That is a categorical recovery (0 → 1), independent of scale or engine, and it is the honest core of CH's value—not a token percentage.
CH does not make a small interruption cheaper or more correct with a strong model—the tree carries enough signal to reconstruct simple state. CH earns its keep as (1) a large-session re-scan avoider (scale effect, not shown on a toy) and (2) the only mechanism that preserves the in-flight-agent record a cold restart categorically cannot. Dated 2026-07-03; a single-scenario sample, not a guarantee.