The climb board: data, table, and chart of every attempt, kernel-published

CHANGELOG.md

@@ -24,6 +24,7 @@ Versions follow [SemVer](https://semver.org).
 
 ### Added
 
+- The climb board: `CLIMB.md` (numbers + attempts table), `climb/<benchmark>.json` (the data), and `climb.html` (a chart of the JSON) on the target's `research-log` branch, kernel-published as runs end. Idempotent: rows merge by run id, files are written only on change.
 - SWEEPS: `launch --array K` fans one command out to K jobs, each with
   `SWEEP_INDEX=0..K-1` in its environment, delivered back as K results
   (artifacts under `results/<name>/<i>/`) in one wake. It counts as one

docs/roadmap.md

@@ -160,6 +160,13 @@ open-ended sweep space.
 - [ ] Task selection reads lessons/ + recent runs/ (the loop's research memory)
 - [ ] Periodic distillation pass: raw reports → bounded lessons/<target>.md
 - [ ] Weekly digest aggregates per-run reports; cost ledger; leaderboard history
+- [x] Climb board on the research-log branch: `climb/<benchmark>.json` +
+      `CLIMB.md` + `climb.html`; a Pages site can serve the same files at the
+      public flip
+- [ ] Local-inference roles (big item, 2026-08-29): open-weight author/judge
+      models served on cluster GPUs (vLLM/SGLang behind the hermes provider
+      seam) — stability against API model retirement, and pinned weights make
+      the research program reproducible
 
 ## Cross-cluster (dev plan, 2026-08-29)
 

src/autoresearch/attempt.py

@@ -279,9 +279,12 @@ def _clear_stage(record: RunRecord) -> RunRecord:
     especially `wake_attempts` ride into `in-review`, where in-review follow-up
     servicing reuses `wake_attempts` as its OWN retry cap — so a run that woke
     once would reach review with a shrunk follow-up budget."""
+    # the run's spend survives the wipe: terminal reporting (the climb
+    # board) reads it after the transition
+    kept = {k: record.stage[k] for k in ("gpu_hours_used",) if record.stage and k in record.stage}
     return dc_replace(
         record,
-        stage={},
+        stage=kept,
         experiment_job_id="",
         deadline=0.0,
         terminal_seen=0.0,