Task #603: P3' write decomposition pipeline

.claude/agents/analyzer.md

@@ -293,6 +293,7 @@ Write first to a local file `.claude/cache/experiment-<N>-clean-result.md` (thro
    - **`**Artifacts:**`** — links to training data, model checkpoints, eval JSONs, figure source, raw completions. The training-data dropdown lives under `### What I ran`; eval examples live near the finding that consumed them (NOT here); this Artifacts block just lists the full artifact links. **GROUND every path-specific artifact claim in a live Hub listing — never type it from the plan's intent.** When you write a bullet that names specific subfolders, checkpoint directories, intermediate-fraction adapters, file counts, or HF Hub paths (e.g. "per-cell LoRA adapters at intermediate fractions {0.25, 0.50, 0.75, 1.00} uploaded to `adapters/issue_<N>/<cell>/`", "520 files at `<path>`"), run `huggingface_hub.list_repo_files` on the relevant repo + revision at write time and copy what the listing actually shows. The `hf` CLI has no `api` subcommand and false-reports "0 files" on a path that exists, so use the Python Hub API (see `.claude/rules/upload-policy.md` for the canonical snippet). If a planned subfolder is missing — e.g. a band-stop callback halted training before the planned intermediate-fraction checkpoint was saved — the body says what is ACTUALLY on the Hub, not what the plan intended; the missing piece becomes a methodology-correction beat inside the relevant `#### <finding>` H4 (the silent-fail rule in CLAUDE.md § "After Every Experiment" #8). A plan-intent claim that doesn't survive the listing is a data-integrity bug that propagates: it gets carried into follow-up-proposer's reuse premises (incident #530→#534, 2026-06-09) and into any future task whose planner mines this body for prior-art artifacts. **Reuse provenance — when ANY reader-facing claim in this body rests on a trained artifact REUSED from a prior issue** (a LoRA adapter, merged checkpoint, training-mix dataset, raw-completion bucket, or `eval_results/` JSON produced by a previous `/issue` run rather than freshly produced by THIS task), record one bullet per reused artifact under this block stating: (a) the producing issue number `#M` as a markdown link to `https://eps.superkaiba.com/tasks/M`; (b) the permanent HF Hub path (pinned to `/tree/<sha>` or `@<sha>`) or repo-relative `eval_results/issue_M/...` path the artifact was pulled from; and (c) a one-line fitness rationale stating WHY this artifact was the right one to reuse — recipe match (same base model + training-recipe / hyperparameters the new question demands), measurement-regime fit (the artifact's eval surface contains the conditions THIS result reads off; for marker work specifically, the artifact is NOT saturated where this read needs headroom — source `log P − base ∈ [5,12]` nat per `.claude/rules/marker-training-recipe.md`), and required conditions present. Mirror the positive fitness check the planner ran at plan §5 / §10 (CLAUDE.md § "Reuse existing trained artifacts when fit-for-purpose — never reuse a wrong one") so the clean-result carries the same justification forward. Format: `- Reused <kind> from [#M](...): <hf path or local path> — fit: <one line: recipe + regime + conditions>`. Source the reuse list from the plan body (§5 reusable + §10/§11 artifact citations) and from any explicit `Source: #M` / `from-issue` references in the training-script SHA at the `**Code:**` link; never invent reuse the plan didn't approve. When THIS task produced every artifact it stands on, omit the reuse-provenance bullets entirely (most fresh-train experiments). The clean-result-critic Lens 5 audits this.
    - **`**Compute:**`** — wall time, GPU type/count, pod label.
    - **`**Code:**`** — dataset-build script, pipeline driver, Hydra config, git commit hash, one-block reproduce snippet.
+   - **`**Context:**`** — run-context provenance (REQUIRED for v2 bodies; SPEC.md § `**Context:**` row; verifier check 17). Three bullets: **Created / run** (frontmatter `created_at` + the date/window results landed), **Follow-up to** (`[#K](https://eps.superkaiba.com/tasks/K) — <one line>` from frontmatter `parent_id` / the lineage that seeded the task, or `fresh direction (no parent)`; for same-issue follow-up rounds also name the round's `followup_label`), and **Originating prompt(s), verbatim** (blockquoted). Source the prompt from, in priority order: frontmatter `origin_prompt`; the ORIGINAL task body's `## Provenance` section — read it BEFORE Step 6's `set-body --snapshot` replaces the live body (post-promotion it lives only in `original-body.md`, so on re-drafts read it from there); and `epm:followup-scope v1` markers with `source: user-chat` (via `task.py latest-marker` / `view --json`, never a hand-built `tasks/...` path). VERBATIM means verbatim — never paraphrase, trim, or fix typos. When no prompt was recorded, write the literal `origin prompt not recorded` — never omit the row, never fabricate. Provenance lives ONLY here: the "state facts, not sources" rule still bans prompt/person attributions in `## TL;DR` and finding prose.
 
    **Confidence lives in the H1 title tag only** for v2 nested-design bodies. Do NOT emit a `Confidence: LOW|MODERATE|HIGH — <rationale>` sentence in `## Reproducibility`. There is NO "Why confidence is where it is" section. If you need to convey the binding constraint that drove the title's confidence level, weave it into the relevant `#### <finding>` read paragraph.
 
@@ -434,7 +435,7 @@ uv run python "$REPO_ROOT"/scripts/verify_task_body.py --file .claude/cache/expe
 
 Every FAIL must be fixed before posting. WARNs may ship when explicitly acknowledged in the body (e.g. the qualitative-data-link WARN for runs whose raw completions weren't uploaded — pair with a "re-run with raw-completion upload" bullet in Next steps). Do NOT proceed to Step 6 until the verifier is FAIL-free.
 
-The verifier enforces the mechanical checks for the 2-content-section nested-design (v2) spec (see `scripts/verify_task_body.py` docstring for the canonical enumeration): body-nonstub (check 0, defense against the cache → body.md silent-handoff failure); no-duplicate-frontmatter (check 0b); title confidence tag (`(LOW|MODERATE|HIGH confidence)`); three required H2s in order (`## Human TL;DR`, `## TL;DR`, `## Reproducibility`) — a stray `## Details` or `## Figure` H2 is a hard FAIL (forces clean migration to the 2-content-section spec); `## TL;DR` opens with the Motivation block, either as `### Motivation` H3 (preferred) or `**Motivation:**` boldface bullet (legacy form); v2 nested-shape check (sentinel-gated) — `### Motivation` / `### What I ran` / `### Findings` H3s in order with ≥1 `#### <finding>` H4 child under `### Findings`; at least one `![alt](url)` image exists inline under `## TL;DR` (every finding H4 carries its own figure); every image URL is absolute + commit-pinned; Confidence — for v2 bodies the H1 title tag is the source of truth (PASSes with NO body sentence); for legacy (pre-sentinel) bodies `Confidence: …` line scanned whole-body, matches the title's level + ≥20 chars of rationale (lives in `## Reproducibility` by convention); `## Reproducibility` carries all three boldface subgroups (`**Artifacts:**`, `**Compute:**`, `**Code:**`); URL permanence in Reproducibility (HF Hub `/tree/<ref>`, WandB `/runs/<id>`, GitHub `/blob/<sha>`; no `main`/`master`/`HEAD`); no `{{` / `TBD` / `see config` / `default` sentinels in Reproducibility (write `n/a` explicitly); cherry-picked label preceding every sample-output block in `## TL;DR` (fenced OR `<details>`-wrapped table/long-text); qualitative-data link in the same prelude or inside the `<details>` block (raw text-level artifact, not aggregate); Reproducibility lr matches plan (check 16, v2-only) — the learning rate in the Parameters table must appear in the approved `plans/plan.md` (FAIL unless a documented run-vs-plan deviation downgrades it to WARN; NO-OP PASS when it cannot reconcile). Soft WARN: `check_details_narrative_flow` flags outline-label H3s + figure-dump runs (>2 consecutive figures without prose between) inside `## TL;DR` (but does NOT flag `### Findings` or `### What I ran` — those are required structural H3s under v2). See `CLAUDE.md § Experiment Report Structure` for the canonical body shape this verifier checks.
+The verifier enforces the mechanical checks for the 2-content-section nested-design (v2) spec (see `scripts/verify_task_body.py` docstring for the canonical enumeration): body-nonstub (check 0, defense against the cache → body.md silent-handoff failure); no-duplicate-frontmatter (check 0b); title confidence tag (`(LOW|MODERATE|HIGH confidence)`); three required H2s in order (`## Human TL;DR`, `## TL;DR`, `## Reproducibility`) — a stray `## Details` or `## Figure` H2 is a hard FAIL (forces clean migration to the 2-content-section spec); `## TL;DR` opens with the Motivation block, either as `### Motivation` H3 (preferred) or `**Motivation:**` boldface bullet (legacy form); v2 nested-shape check (sentinel-gated) — `### Motivation` / `### What I ran` / `### Findings` H3s in order with ≥1 `#### <finding>` H4 child under `### Findings`; at least one `![alt](url)` image exists inline under `## TL;DR` (every finding H4 carries its own figure); every image URL is absolute + commit-pinned; Confidence — for v2 bodies the H1 title tag is the source of truth (PASSes with NO body sentence); for legacy (pre-sentinel) bodies `Confidence: …` line scanned whole-body, matches the title's level + ≥20 chars of rationale (lives in `## Reproducibility` by convention); `## Reproducibility` carries all three boldface subgroups (`**Artifacts:**`, `**Compute:**`, `**Code:**`); URL permanence in Reproducibility (HF Hub `/tree/<ref>`, WandB `/runs/<id>`, GitHub `/blob/<sha>`; no `main`/`master`/`HEAD`); no `{{` / `TBD` / `see config` / `default` sentinels in Reproducibility (write `n/a` explicitly); cherry-picked label preceding every sample-output block in `## TL;DR` (fenced OR `<details>`-wrapped table/long-text); qualitative-data link in the same prelude or inside the `<details>` block (raw text-level artifact, not aggregate); Reproducibility lr matches plan (check 16, v2-only) — the learning rate in the Parameters table must appear in the approved `plans/plan.md` (FAIL unless a documented run-vs-plan deviation downgrades it to WARN; NO-OP PASS when it cannot reconcile); Reproducibility Context provenance row (check 17, v2-only) — the `**Context:**` row (created/run dates, follow-up lineage, verbatim originating prompt) must be present (FAIL only when recorded origin data — frontmatter `origin_prompt` or a `## Provenance` section in `original-body.md` — exists but the body dropped it; WARN otherwise). Soft WARN: `check_details_narrative_flow` flags outline-label H3s + figure-dump runs (>2 consecutive figures without prose between) inside `## TL;DR` (but does NOT flag `### Findings` or `### What I ran` — those are required structural H3s under v2). See `CLAUDE.md § Experiment Report Structure` for the canonical body shape this verifier checks.
 
 ### Step 6: Promote the source experiment to a clean-result (inline)
 

.claude/agents/clean-result-critic.md

@@ -101,6 +101,12 @@ anti-pattern audit:
 #        in the Parameters table must appear in `plans/plan.md`
 #        (FAIL unless a documented run-vs-plan deviation → WARN; NO-OP
 #        PASS when it cannot reconcile). Task #489's 1e-4-vs-2e-6 typo.
+#   17. Reproducibility Context provenance row (v2-only) — the
+#        `**Context:**` row ships created/run dates, follow-up lineage,
+#        and the verbatim originating prompt (FAIL only when recorded
+#        origin data — frontmatter `origin_prompt` or original-body.md
+#        `## Provenance` — exists but the body dropped it; WARN
+#        otherwise; legacy bodies skip).
 #   13. (WARN) TL;DR narrative flow — outline-label H3s + figure-dumps
 uv run python scripts/verify_task_body.py --issue <N>
 
@@ -119,9 +125,13 @@ the verifier's FAILs into two classes before deciding the verdict:
   `![alt](url)` figure exists anywhere under `## TL;DR` (check 4), a
   Reproducibility boldface subgroup is absent (check 7), a retired
   `## Details` / `## Figure` H2 is present (check 2 clean-migration),
-  the body is a stub (nonstub check), or the Reproducibility learning
+  the body is a stub (nonstub check), the Reproducibility learning
   rate does not match the plan (check 16) — a wrong load-bearing
-  hyperparameter is a data-integrity defect, never cosmetic. These are
+  hyperparameter is a data-integrity defect, never cosmetic — or
+  recorded origin provenance was dropped (check 17 FAIL: frontmatter
+  `origin_prompt` / an original-body `## Provenance` section exists but
+  the body carries no `**Context:**` row; the check's WARN form — no
+  recorded origin data — is not a FAIL and never blocks). These are
   like a missing/wrong report section: record the failed check as a
   blocking finding, but STILL read all fifteen lenses in the same
   pass and report every substantive finding you see. **Beyond the
@@ -453,6 +463,25 @@ in Lens 12. Lens number kept stable so downstream tooling that reads
   been linked.
 - No `{{`, `TBD`, `default`, `see config` sentinels — write `n/a`
   explicitly when truly non-applicable (verifier check #9).
+- **Context-row audit (run-context provenance; v2 bodies).** The
+  `**Context:**` row in `## Reproducibility` (SPEC.md
+  § `**Context:**` row; verifier check 17 covers presence — this
+  bullet adds the substantive read) must carry: (a) **real dates** —
+  the created date matches frontmatter `created_at`, the run
+  date/window is plausible against the events.jsonl timeline; (b)
+  **correct lineage** — the `Follow-up to` line matches frontmatter
+  `parent_id` / the Motivation's actual prior-task citation (a
+  fabricated or wrong parent is a FAIL), or says `fresh direction
+  (no parent)`; (c) **verbatim prompts** — cross-check the quoted
+  originating prompt against frontmatter `origin_prompt` and/or the
+  `## Provenance` section in `original-body.md`; a paraphrased,
+  trimmed, or typo-corrected prompt is a FAIL (verbatim means
+  verbatim), and the literal `origin prompt not recorded` is
+  accepted only when no origin data actually exists. Also confirm
+  provenance stays CONFINED to this row — prompt/person attributions
+  woven into `## TL;DR` or finding prose violate the "state facts,
+  not sources" rule. Forward-only: legacy (pre-sentinel) bodies are
+  never failed for lacking the row.
 - **Reuse-provenance audit (semantic, not mechanical).** When any
   reader-facing claim in `## TL;DR` rests on a trained artifact
   REUSED from a prior issue — a LoRA adapter, merged checkpoint,