fix: evolution formula must commit results via PR before closing (#1047)

- Add `cleanup` step: removes per-generation candidate files and
  generation_*.jsonl records after they are aggregated into the evidence
  file, preventing disk exhaustion (cf. run #1025 at 91% usage).

- Rewrite `deliver` step with mandatory ordering:
  1. `git checkout -- .` to discard unrelated working-tree modifications
     before staging result files (evidence JSON, champion .push3, manifest).
  2. Commit to branch `evidence/evolution-run-{run_id}` (not directly to main).
  3. Push and create PR — if this fails, post an error comment and leave the
     issue OPEN; do not proceed to step 4.
  4. Post summary comment only after PR URL is confirmed, with mandatory
     link to the PR.

- Update `products.evidence_file` delivery to PR branch (was "commit to main").
- Update `products.issue_comment` to enforce ordering and non-close-on-failure.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

formulas/run-evolution.toml

@@ -196,36 +196,83 @@ the run; "no_improvement" otherwise.
 output = "evidence/evolution/{date}.json"
 schema = "evidence/README.md"
 
+[[steps]]
+id          = "cleanup"
+description = """
+Remove intermediate per-generation candidate files that are not part of the
+final results.  Only the following files are retained after this step:
+
+  tmp/evolution/run_NNN/best.push3          global champion
+  tmp/evolution/run_NNN/diff.txt            constant delta vs seed
+  tmp/evolution/run_NNN/evolution.log       full run transcript
+  tools/push3-evolution/seeds/run{NNN}_*.push3
+                                            top-N newly admitted seeds
+                                            (≤ elites per generation)
+
+Files removed:
+  tmp/evolution/run_NNN/generation_*.jsonl  per-candidate fitness records
+                                            (already aggregated into evidence)
+  tmp/evolution/run_NNN/candidate_*.push3   intermediate per-generation
+                                            candidates that are not elites
+
+Rationale: the evolution box reached 91% disk utilisation in run #1025 because
+these intermediate files were never cleaned up.  Aggregated fitness data is
+preserved in evidence/evolution/{date}.json; the per-candidate .push3 files for
+non-elite generations are not needed once the evidence file is written.
+"""
+
 [[steps]]
 id          = "deliver"
 description = """
-Commit evidence and champion files, open PR, post summary comment.
+Commit results to a branch, push, open PR, then post summary comment.
+ORDERING IS MANDATORY — each sub-step must complete before the next begins.
+Do NOT post to the issue before the PR URL is available.
 
-1. Commit evidence/evolution/{date}.json to main.
-
-2. Open a Codeberg PR targeting main containing:
+1. CLEAN GIT STATE
+   Run `git checkout -- .` to discard any working-tree modifications that are
+   NOT part of the evolution results (e.g. .sol files left over from a prior
+   session, scratch files).  Only stage files that belong to this run:
+     - evidence/evolution/{date}.json
      - tools/push3-evolution/seeds/evo_run{NNN}_champion.push3
-       (copied from tmp/evolution/run_NNN/best.push3)
-     - Updated tools/push3-evolution/seeds/manifest.jsonl
-       (with newly admitted candidates and fresh fitness scores)
-   PR title: "evo: run{NNN} champion — fitness={best_fitness}"
-   PR body: generation-by-generation table (gen, best, mean, worst fitness),
+     - tools/push3-evolution/seeds/manifest.jsonl
+   Verify `git diff --check` passes before committing.
+
+2. COMMIT TO BRANCH
+   Create branch evidence/evolution-run-{run_id} from master.
+   Commit the staged result files with message:
+     "evo: run{NNN} results — fitness={best_fitness}"
+   The commit MUST include all three files above.
+
+3. PUSH AND CREATE PR
+   Push the branch to origin.
+   Open a Codeberg PR targeting master:
+     Title: "evo: run{NNN} champion — fitness={best_fitness}"
+     Body:  generation-by-generation table (gen, best, mean, worst fitness),
             top-3 admitted candidates with fitness scores, constant diff vs
             seed (from diff.txt), link to evidence file.
+   If `git push` or PR creation fails:
+     a. Post an error comment to the originating issue with the failure reason
+        and the path of the local evidence file.
+     b. Leave the issue OPEN.
+     c. Exit with a non-zero status — do NOT proceed to step 4.
 
-3. Post summary comment to originating issue:
+4. POST SUMMARY COMMENT (only after PR URL is confirmed)
+   Post a comment to the originating issue containing:
      - Verdict (improved / no_improvement).
      - Best fitness achieved and which generation it was found in.
      - Admission count: N candidates added to seed pool.
-     - Link to champion PR (if new best found) and evidence file.
-     - If no_improvement: include best fitness achieved and seed pool size.
+     - Link to the champion PR (required — do not post without it).
+     - Link to evidence file committed in the PR.
+     - If no_improvement: best fitness achieved and seed pool size.
+   Do NOT close the issue in this step; closing is the orchestrator's
+   responsibility once the PR is merged.
 """
 
 # ── Products ───────────────────────────────────────────────────────────────────
 
 [products.evidence_file]
 path     = "evidence/evolution/{date}.json"
-delivery = "commit to main"
+delivery = "PR to main (same PR as champion_files, on branch evidence/evolution-run-{run_id})"
 schema   = "evidence/README.md"   # see ## Schema: evolution/YYYY-MM-DD.json
 
 [products.champion_files]
@@ -240,9 +287,11 @@ delivery = "PR to main (same PR as champion_files)"
 note     = "Updated with newly admitted entries and fitness scores from evaluate-seeds."
 
 [products.issue_comment]
-delivery   = "post to originating issue"
-content    = "verdict (improved/no_improvement), best fitness, generation found, admission count, link to champion PR and evidence file"
-on_failure = "include best fitness achieved, last generation completed, full log available in tmp/evolution/run_NNN/evolution.log"
+delivery         = "post to originating issue AFTER PR is created and URL is confirmed"
+content          = "verdict (improved/no_improvement), best fitness, generation found, admission count, link to champion PR (mandatory), link to evidence file"
+on_pr_failure    = "post error comment with failure reason and local evidence path; leave issue OPEN; do not close"
+on_run_failure   = "include best fitness achieved, last generation completed, full log available in tmp/evolution/run_NNN/evolution.log; do not close issue"
+ordering_note    = "The comment MUST NOT be posted before the PR URL exists.  Closing the issue is the orchestrator's responsibility after PR merge, not this formula's."
 
 # ── Resources ──────────────────────────────────────────────────────────────────