shapeup-sdlc 1.7.0 → 3.0.1

package/skills/tech-lead/references/protocol.md

@@ -0,0 +1,832 @@
+# Protocol — the round loop, the delegation shape, and the invariants behind both
+
+Three things that only make sense together: what a round DOES, how each step is dispatched, and
+which rules the code enforces so a model does not have to remember them.
+
+On a spec with committed scope contracts all of this is CODE —
+`skills/tech-lead/workflows/shapeup-run.js`. This file is the readable rationale for why that code
+is shaped the way it is, plus the verbatim prose path for the lanes the script does not cover
+(`--tiny`, and any spec with no `scopes/*.md` yet).
+
+---
+
+# Part 1 — The round loop
+
+The orchestration loop in detail. A "round" is one BUILD phase followed by exactly one
+EVAL phase. The feature is done when an EVAL round returns PASS.
+
+> `tasks/_index.md` referenced throughout this file lives in the LOCAL gitignored root
+> (`.shapeup/<slug>/tasks/`, v3.2), not the
+> SHARED spec dir. See tech-lead SKILL.md GATE L1b for the bootstrap step that regenerates it
+> when missing.
+
+**On a spec with committed scope contracts, this loop is now CODE, not prose** —
+`skills/tech-lead/workflows/shapeup-run.js`'s own `while (round <= args.budgets.maxRounds)` loop
+implements exactly the shape below (PASS → ship; `max_rounds` exceeded → GATE H `breaker: outer`;
+else the next round builds bugs only), with the EVAL-timing rule, the regression rule, and the
+three-level breaker all enforced as branches in that script rather than described here for a model
+to follow. Read that script's own comments for the mechanics; this file keeps the historical
+rationale and the parts of the protocol the script does not cover (ESCALATE adjudication,
+discovered-task reconciliation mid-BUILD — see its banner
+for the full list) — and, unchanged, the loop below for a `--tiny` run or a spec with no scope
+contracts, which `shapeup-run.js` is out of scope for by design.
+
+```
+round r = 1
+loop:
+    BUILD(r)                          # see r=1 vs r>1 below
+    assert board 100% done            # GATE L2 — advisory: the hook warns, you decide
+    verdict, bugs = EVAL(r)           # ONE spec-evaluator --feature pass
+    if verdict == PASS:
+        SHIP; break
+    if r >= max_rounds:
+        ESCALATE(bugs); break         # honest stop — no infinite loop
+    r = r + 1                         # next round builds bugs only
+```
+
+## BUILD(r) semantics
+
+| | r = 1 (initial build) | r > 1 (fix build) |
+|--|----------------------|-------------------|
+| Input | the whole task board | the bug list from EVAL(r-1) |
+| Scope | every ready task, dependency/layer order, until board all ✅ | only the tasks/areas named by bugs |
+| Command | compile-order `--next` → task-executor `--order` → ingest, looped | compile-order `--task <id> --operation fix` → dispatch → ingest, per bug |
+| Passing areas | n/a | never touched |
+| SPIKEs | resolved first (they block) | only if a bug is a SPIKE finding |
+
+Re-opening tasks in r>1: the fix order's WorkResult reports the task `partial` while failing
+and `done` when re-verified; ingest-result flips the board accordingly. The board reflects the
+churn so the next EVAL sees a green board again.
+
+Discovered Tasks:
+If WorkResults carry `discoveries[]` during BUILD, ingest-result appends them to the discovery
+ledger (`.shapeup/<slug>/discovery/ledger.md`) and the build loop pauses after the current
+tasks are done. Compile + dispatch a reconcile order (ba-pitch-analyzer, operation: reconcile).
+This reconciles them into new tasks and invariants and updates the board; the tech lead bumps
+`discovered_rounds` in harness-run.md, then routes back to GATE L1b (Board Review) for PO
+approval of the new tasks and estimates before resuming the BUILD loop.
+
+## The EVAL timing rule (the core constraint)
+
+EVAL fires **once** per round and **only** when GATE L2 has confirmed the board is 100% done.
+It is never:
+- called per task,
+- called inside the BUILD loop,
+- called on a partial board.
+
+## Regression rule (r > 1) — QA-meeting Bước 1c
+
+A fix round changes code; a fix can break what passed. Therefore EVAL(r>1) scope is **not**
+just the fixed bugs:
+
+```
+EVAL(r) for r > 1:
+  touched_UCs = every UC referenced (use_case_refs) by a task re-opened in BUILD(r)
+  scope = fixed bugs' criteria
+        + FULL re-run of `## Test Surface` rows for every touched UC
+          (test-surface-conformance dimension, when active)
+        + completeness re-check (cheap, static)
+  untouched UCs' surfaces: NOT re-run (their code didn't change; re-probing everything
+  every round would turn cheap end-of-round QA into a full-suite tax)
+```
+
+Pre-v2.9 specs (no Test Surface anywhere): the rule degrades to bug-criteria-only, as
+before — and the verdict report notes `regression coverage: none (no test surface)`.
+Honest reporting over silent coverage claims.
+
+## QA edge hunt (post-PASS, pre-ship)
+
+When EVAL(r) returns PASS for the **first** time in a run, the orchestrator delegates one
+`/qa-edge-hunter` pass before SHIP (skippable via `--no-qa`, same spirit as `--no-eval`).
+QA is a pure worker: no verdict, no score, no gate — it writes `~` findings to
+`discovery/ledger.md` and a `qa/hunt-report.md`. Triage happens at SHIP/GATE L4:
+- all findings stay `~` → SHIP; findings carry over as raw ideas (debt-free).
+- PO/TL promote any to must-have → a fix round r+1 (those items only) → EVAL
+  never a second full hunt) → back to L4.
+- Circuit breaker applies: out of rounds/appetite → ship with `~` findings recorded.
+QA never runs on a FAIL round — a build that hasn't passed conformance isn't worth
+edge-hunting yet.
+
+Rationale (from the long-running harness work): a single end-of-round QA pass over the
+running feature is cheap relative to the build (minutes vs hours) and catches the
+last-mile defects, whereas grading every task multiplies evaluator cost for little gain
+once the generator is competent. If the build round didn't finish, there is nothing
+coherent to evaluate yet.
+
+## Stop conditions
+1. **PASS** — EVAL(r) verdict is PASS → SHIP.
+2. **max_rounds (OUTER breaker)** — r would exceed `--max-rounds` (default 3) without PASS →
+   ESCALATE: print the residual bug list, the rounds used, and hand the decision to the PO
+   (scope contracts present: also `/scope-hammer --breaker outer`). Do not start another
+   build round automatically.
+3. **attempt_budget (INNER breaker, scope contracts only)** — a single scope's T0 attempt
+   loop exhausts `--attempts` (default 5) without reaching a trial that is both `kept` and
+   T0-green, **or** `no_progress_k` consecutive trials come back non-`kept` (the stagnation
+   term, default 2 — `harness compile` prints it as a JSON breaker object on stderr) →
+   does NOT stop the round; queues a hammer PROPOSAL for GATE H and moves to the next scope
+   in sequence. See "Three-level circuit breaker" below.
+4. **wall_clock_budget (DEADLINE breaker, opt-in)** — elapsed seconds since the run receipt
+   exceed `--wall-clock-budget` → do NOT start another build round or another scope; go
+   straight to GATE H (`/scope-hammer --breaker deadline`). Checked with
+   `harness verify budget` at every round boundary and enforced by `harness verify budget`,
+   which denies a `task-executor` dispatch past the deadline while leaving `spec-evaluator`,
+   `scope-hammer` and `qa-edge-hunter` reachable — a run past its deadline
+   must still be able to judge, hammer, and close. Off unless configured.
+5. **Hard error** — a sub-skill fails irrecoverably (e.g. spec folder gone, app won't
+   build at all) → stop and report; do not retry blindly.
+6. **User halt** — at any L-gate the user can stop the run; the ledger preserves state for
+   `--from` resume.
+
+## Three-level circuit breaker
+
+```
+OUTER    round_budget (max_rounds)  — the six-week-timebox analog. Decremented once per
+                                       round at GATE L2, regardless of how many scopes it
+                                       covered. Hitting 0 → GATE H immediately (§ above).
+INNER    attempt_budget (per scope) — decremented once per T0 attempt inside BUILD round r.
+                                       Hitting its cap WITHOUT a T0-green result trips the
+                                       inner breaker for that scope only: the scope is queued
+                                       as a hammer PROPOSAL (not a hard stop) and the round
+                                       moves on to the next scope in the L1b sequence.
+                                       no_progress_k (default 2) is COMPOSED INTO this same
+                                       breaker rather than added beside it as a fourth budget:
+                                       attempt_budget counts attempts and cannot see that the
+                                       last two produced nothing, so k consecutive non-`kept`
+                                       trials queue the same GATE H proposal early.
+DEADLINE wall_clock_budget_s        — elapsed seconds since the run receipt. Opt-in; off
+                                       unless set at L0. Tripping routes to GATE H with
+                                       --breaker deadline. Enforced by `harness verify budget`.
+```
+
+**Why the third one exists — and it corrects an earlier diagnosis.** A run killed at an external
+time cap looks like a stall from outside: no verdict, nothing to show. The natural reading is that
+it hung at a gate. Often the transcript says otherwise — steady turns, steady writes, gate markers
+advancing, zero stall signals. The run was working when the clock ran out.
+
+Both existing breakers count *events*, not time: `round_budget` moves once per round,
+`attempt_budget` once per T0 attempt. Neither can observe that round 1 has been running for
+twenty-nine minutes, so a run can burn its entire wall clock with both breakers untouched. The
+cost is not the missing verdict — it is that a run killed from *outside* ships nothing, not even the
+scopes that were already green. A breaker that trips from the inside routes to GATE H, where
+scope-hammer compares the shippable subset against the baseline and ships what works. Same clock,
+different ending.
+Nesting rationale: a struggling scope should not freeze every other scope's progress
+in the same round — only running out of *rounds* (the real six-week analog) stops the whole
+run. A scope that trips its inner breaker still gets judged fairly at GATE H: scope-hammer
+compares "ship without this scope" against the baseline, same as any other cut candidate — it
+is never silently dropped, and it is never allowed to block scopes that ARE working.
+
+## Isolated attempt loop — one T0 attempt, in detail (scope contracts only)
+
+**This is `shapeup-run.js`'s inner per-scope loop, as code**, with one addition the prose below
+never had: before opening a scope's attempt loop, the script checks whether THIS round already
+has a green T0 verdict for that scope on disk, and skips it when it does — the resumability a
+mid-BUILD kill needs, that a session narrating this loop from memory could never guarantee.
+
+Whichever runs it, the loop is a **ratchet**: every attempt is scored against the last kept one,
+and the working tree moves forward only when the score strictly improves. `harness verify t0` makes
+that decision and acts on the tree itself, inside the runtime — its caller reads the decision, it
+never makes it.
+
+```
+compile-order --scope … --round N --attempt M
+                                   → zero-memory WorkOrder (scope contract + this scope's
+                                      tasks + digested errors + ledger decisions +
+                                      trial_history: the last 8 trials for this scope, each
+                                      with score, status, delta and top-3 digest, CROSSING the
+                                      round boundary — compiled facts, no chat history by
+                                      construction)
+dispatch task-executor --order …   → code within substrate; WorkResult in results/
+ingest-result <result>             → board/ledger writes
+(a worker cannot escalate)         → WorkResult carries no escalates field. A phase that
+                                      cannot finish leaves no artifact, the post-condition
+                                      fails, and the run ABORTS naming the phase. Resolve it
+                                      yourself and record the answer in round-ledger.md, which
+                                      the NEXT attempt's fresh context reads back.
+harness verify t0                      → fixtures + DB probe + (on green) seesaw, then scores the
+                                      attempt against the baseline trial and snapshots or
+                                      restores the tree. Branch on `status` from its stdout
+                                      JSON — the tree action has ALREADY happened:
+  kept      strictly better, INCLUDING red-but-improved (2/5 → 4/5 fixtures — the whole point
+              of the ratchet). Tree snapshotted to refs/shapeup/<scope_id>/kept.
+              overall=green → the attempt loop breaks; scope reaches DOWNHILL_EXECUTION.
+              Still red → loop, and attempt M+1 now builds ON attempt M.
+  reverted  not better. A tie on a RED score is not better either — but a tie on a GREEN one
+              KEEPS, because a spec-conformance fix cannot raise a score that is already at
+              full marks, and reverting it would discard exactly the work a fix round exists
+              to do. Tree already restored from the last kept
+              snapshot. Subsumes the retired stash-and-retry branch: a FINISHED scope's broken
+              fixture raises score.regressions and reverts through this same rule, which
+              is why seesaw runs before anything is declared green.
+              rather than the code. Tree kept, baseline reset. Not a verdict, not a failure.
+  crash     a fixture command failed to spawn or timed out; tree restored. Fix the fixture,
+              not the code.
+  (on any red, `discovered_tasks` carries the AEGIS {file, line, core_message} triples, which
+   compile-order folds into the NEXT attempt's order as digested_errors — no separate dispatch)
+```
+
+**The exit code is not the branch selector.** `harness verify t0` exits 0 on T0-green and 1 on
+T0-red (2 on bad argv), mirroring the `oracles/*` convention — so a `kept` red-but-improved
+attempt, the exact case the ratchet exists for, exits 1. Branch on `status` from the stdout
+JSON; never on `$?`, and never wire this call into a `set -e` / `&&` chain that would read a
+non-zero exit as "stop".
+
+This replaces the old flat per-task loop for any scope that has a contract;
+scopes/specs without one keep the v0.2.6 behavior verbatim (see BUILD(r) table above).
+
+## --no-eval (skip evaluation)
+A tech-lead judgment, surfaced at GATE L2: if the feature is clearly within what the model
+builds reliably solo, the evaluator is optional overhead. With `--no-eval`, after GATE L2
+the run goes straight to SHIP with verdict `not-evaluated` recorded in the ledger and a
+clear note that nothing was verified beyond task-executor's own per-AC evidence checks.
+
+## Round-cost intuition
+Build dominates; eval is cheap. Expect each EVAL round to cost a small fraction of a BUILD
+round. This is why running eval once per round (not per task) is the right trade: you pay a
+little QA at the end of each build and keep the expensive build coherent in between.
+
+<!-- test requirement -->
+kept|reverted|rebased|crash, decided in harness verify t0 decideStatus()
+
+---
+
+# Part 2 — Delegation
+
+The tech lead invokes the build-phase skills and reads their handoff files. It never
+reimplements their logic. Each sub-skill keeps its own gates; pass `--auto` to a sub-skill
+only when the run's auto level is `--auto` or `--unattended`.
+
+## Invocation mechanism — Agent, not Skill
+
+Every "Invoke:" line below means: call the **Agent** tool — a real subagent, on its own
+context and (where a role is named) its own model — whose prompt tells it to run
+`Skill(shapeup-sdlc-plugin:<name>)` with the given args and report back the artifacts. It does
+**not** mean the tech lead calls the `Skill` tool itself. A direct `Skill` call executes
+inline, in the tech lead's own turn, on the tech lead's own model — that silently drops GATE
+L0.8's model matrix (there is no per-role model left to route once the call is inline) and
+breaks the isolation the zero-memory-handoff design (`protocol.md`, the compiled
+WorkOrder in `orders/r<N>-a<M>.json`) already assumes every worker below has.
+
+Standard shape (pure-skill architecture v1.0 — the envelope port):
+```
+1. node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" compile <mode flags>      # → orders/<id>.json
+2. Agent({
+     description: "<short task description>",
+     subagent_type: "general-purpose",
+     model: "<role model resolved at GATE L0.8>",
+     prompt: "Call Skill(shapeup-sdlc-plugin:<skill>) --order <orders/<id>.json>.
+              Report back: the WorkResult path (.shapeup/<slug>/results/<id>.json)."
+   })
+3. node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" reduce ingest .shapeup/<slug>/results/<id>.json
+```
+The WorkOrder carries everything the worker may rely on (payload, decisions, digested errors,
+substrate write-contract); the WorkResult carries everything the worker used to write into
+shared files. `harness verify envelope` runs as a PreToolUse hook on Skill|Agent and DENIES a
+dispatch whose `--order` file is missing or schema-invalid. Workers write only their own
+domain artifacts inside their substrate — never boards, ledgers, or run-state.
+
+Role → model, resolved once at GATE L0.8 from the `orch`/`exec`/`eval`/`qa` matrix
+(`harness verify t0` is mechanical tooling run directly via Bash, never an Agent — zero LLM
+tokens):
+
+| Skill | L0.8 role | Why this tier |
+|-------|-----------|---------------|
+| translator | exec | one-shot text transform — builder tier |
+| orient (Scout) | exec | reads/spikes code — builder tier, not judgment |
+| ba-pitch-analyzer | exec | planner — builder tier, not judgment |
+| scope-architect | exec | scope-contract author (sole writer of scopes/*.md) — builder tier |
+| task-executor | exec | the builder itself |
+| spec-evaluator | eval | the single judge (judge ≠ doer) — keep its own matrix key even if a PO points it at the same model as `exec`, so it can be split later without a harness change |
+| qa-edge-hunter | qa | cheapest tier by design — exploratory breadth over depth |
+| scope-hammer | exec | census + baseline comparison, proposes only — not a verdict |
+| coach | exec | categorization gate, not a verdict |
+
+The tech lead itself is `orch` — this conversation, never delegated to.
+
+The order is **Orient (7) → Map Scopes (8) → Build (9) → Eval**, faithful to Shape Up: the
+team orients before any board exists, so the planner's board is reality-born. The tech lead
+is the **sole writer of run-state** (`harness-run.md`) — it passes each worker the run
+metadata it needs (`feature`, `spec`, `stack`, `discovered_rounds`, `--auto`) as **args**;
+workers keep only their own product-idempotency key and emit domain artifacts.
+
+## 0. LANGUAGE GATE → translator (GATE L0, only if non-English)
+```
+Invoke via Agent (model: exec): Skill(shapeup-sdlc-plugin:translator) --check "<intake path>"
+                 # detect-only, writes nothing
+  English      → skip; ORIENT against the original.
+  non-English  → Agent (model: exec): Skill(shapeup-sdlc-plugin:translator) "<intake path>" [--auto]
+                 # full pass
+                 Writes: <name>.en.md (English copy; original untouched) + glossary.md
+                         + translation-report.md.
+                 ORIENT against the <name>.en.md copy.
+Read back: the detect table (--check) / the .en.md path + residual scan result (full pass).
+Authority: translator normalizes language only — it does not orient/plan/build/judge. The tech
+lead never translates itself; it only detects and sequences this step before ORIENT.
+```
+
+## 1. ORIENT → orient (the Scout, step 7) — runs BEFORE planning
+```
+Invoke via Agent (model: exec): Skill(shapeup-sdlc-plugin:orient)
+        --pitch "<kicked-off pitch path>" --spec <path> --stack "<hint>" [--auto]
+Owns:   its own GATE O-A/O-B (or runs straight through under --auto)
+Writes: .shapeup/<slug>/orient/ → code-surface.md, spike-<area>.md, discovered-seed.md, hill-signal.md (LOCAL run-trace)
+Read back: hill-signal.md (render the area-level Hill at GATE L1a) + the spiked area/result.
+Why first: at Orient time NO board exists; the Scout's map + discovered seed make the planner's
+        board reality-born instead of imagined. The four artifacts are the orient→ba contract.
+Authority: pure worker — no code, no board, no run-state, no reporting.
+```
+
+## 2. MAP SCOPES → ba-pitch-analyzer + scope-architect (step 8, orient-informed)
+```
+Order A (the spec tree + board):
+  compile-order --operation analyze --slug <slug> --worker ba-pitch-analyzer
+    --payload '{"pitch": "<path>", "lens": "<lens>", "orient_dir": ".shapeup/<slug>/orient/"}'
+  Agent (model: exec): Skill(shapeup-sdlc-plugin:ba-pitch-analyzer) --order <path>
+  The order hands it code-surface.md (Phase-1 ingest, no re-scan), discovered-seed.md (task
+  gen from reality), spike-<area>.md (feasibility/contracts).
+  Writes (its substrate): spec_folder → _index.md, domain-model.md, ux-behavior.md, usecases/*,
+    contracts/*.contract.md, scope-summary.md (+ api-feasibility.md if third-party) and the
+  Returns: WorkResult (artifacts list + discoveries) → ingest-result.
+Order B (the scope contracts):
+  compile-order --operation map-scopes --slug <slug> --worker scope-architect
+  Agent (model: exec): Skill(shapeup-sdlc-plugin:scope-architect) --order <path>
+  Writes (its substrate): shapeup/<slug>/scopes/*.md + scope-board.md — sole
+    writer. Lint mechanically: node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" verify spec <slug> (PA1/PA2 +
+    substrate disjointness + the spec anchor: SCOPE-ANCHOR/TIER-DIRECTION/SCOPE-DEPS) before GATE L1b.
+Read back: .shapeup/<slug>/tasks/_index.md (the board) + scope-summary.md (Done-when)
+  + the spec-lint verdict.
+Pass-through rule: do not coach the planner to over-specify implementation — keep tech high-level.
+```
+
+## 2b. RECONCILE → ba-pitch-analyzer (discovered task reconciliation, operation: reconcile)
+```
+compile-order --operation reconcile --slug <slug> --worker ba-pitch-analyzer
+  --payload '{"discovered_ledger": ".shapeup/<slug>/discovery/ledger.md"}'
+Agent (model: exec): Skill(shapeup-sdlc-plugin:ba-pitch-analyzer) --order <path>
+Effect: reconciles raw ledger discoveries into board tasks + appended UC invariants/TS rows,
+        inside the reconcile write-contract (frozen zone enforced by the sandbox hook, not
+        prose). Appetite Guard runs mechanically: node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" reduce board.
+Returns: WorkResult → ingest-result (which updates the board and bumps discovered_rounds in
+        harness-run.md — the worker holds no counter).
+Read back: updated tasks/_index.md + scope-summary.md before routing back to GATE L1b.
+```
+
+  Agent (model: exec): Skill(shapeup-sdlc-plugin:ba-pitch-analyzer) --order <path>
+Effect: regenerates the LOCAL task board fresh from the committed usecases/domain-model/scopes
+        — no ledger, no reconciliation. Status bootstraps from committed T0/hill facts at
+        SCOPE granularity; unlocks recomputed by node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" reduce board --write.
+Read back: the freshly regenerated tasks/_index.md before entering BUILD.
+```
+
+## 3. BUILD → task-executor (always through the envelope port)
+```
+r=1 loop:
+  compile-order --next --slug <slug> [--test-cmd "<cmd>"]        # exit 2 = no ready task
+  Agent (model: exec), one fresh subagent per order:
+    Skill(shapeup-sdlc-plugin:task-executor) --order <path>
+  ingest-result <results/<id>.json>   # ticks ACs, marks done, updates board, propagates unblocks
+  Repeat until compile-order --next reports no ready task (board all ✅).
+
+r>1 (fix) per bug:
+  compile-order --task <id> --slug <slug> --operation fix --payload '{"bugs": [<entries>]}'
+  → dispatch + ingest as above. Scope the change to the bug only.
+
+Scope contracts present (isolated attempt loop, per scope, per attempt):
+  compile-order --scope shapeup/<slug>/scopes/<id>.md --round <N> --attempt <M>
+  → orders/r<N>-a<M>.json inlines the scope contract, this scope's tasks, promoted ledger
+    decisions, the previous attempt's AEGIS triples, and `trial_history` — the last 8 trials
+    for this scope (score, status, delta, top-3 digest), CROSSING the round boundary so a fix
+    round cannot re-propose a change the build round already reverted. The zero-memory
+    handoff, compiled from facts only. Dispatch a fresh Agent per attempt (the isolation
+    boundary):
+    Skill(shapeup-sdlc-plugin:task-executor) --order <path>
+  ingest-result — a WorkResult with status "escalated" leaves its artifact unwritten → see 3b below.
+
+Read back: ingest-result's summary line (tasks updated, ACs ticked, unblocked, discoveries) — not raw board
+files. SPIKE tasks close before the tasks they block can build (compile-order enforces the
+dependency order).
+```
+
+## 3b. A worker that cannot finish (scope contracts present, mid-attempt)
+```
+There is NO adjudication dispatch. A worker has no port for "I cannot decide this": WorkResult
+        carries no escalates field, so the question never reaches you as data.
+What you see: the phase produces no artifact. The workflow's post-condition
+        (harness probe resume --require <phase>) fails and the run ABORTS, naming the phase.
+Do: read the phase's result file to find what it could not complete, resolve it yourself —
+        by amending the spec, widening the scope contract, or answering the ambiguity in the
+        round-ledger "Decisions" table — then relaunch. The fast-forward re-dispatches only
+        what is still unfinished.
+Why it aborts rather than pauses: nothing persists an answer between launches, so a pause
+        would relaunch into the same order and hit the same wall. Aborting puts the question
+        in front of a human once instead of looping silently.
+```
+
+## 3c. T0 verify → `harness verify t0` (skill-local; scope contracts present, every attempt)
+```
+Invoke via Bash directly — NOT an Agent, this is deterministic tooling, not a worker:
+  node "${CLAUDE_PLUGIN_ROOT}/kernel/harness.mjs" verify t0 shapeup/<slug>/scopes/<scope-id>.md
+        --round <N> --attempt <M> --seesaw-registry .shapeup/<slug>/seesaw/registry.json
+Effect: runs the scope's e2e fixtures + DB probe, then (on green) the seesaw regression check
+        over every FINISHED scope's fixtures. Writes the verdict artifact spec-evaluator's
+        T0-citation rule will require a citation to, appends one row to t0/trials.jsonl, and — this
+        is the ratchet — scores the attempt against the last kept trial and snapshots or
+        restores the working tree ITSELF. Zero LLM tokens — deterministic tooling, not a
+        judge (this is what keeps "T1 once per round" true even though verification runs
+        every attempt).
+Read back: the stdout JSON — {path, sha256, trial, overall, regression, score, status,
+        baseline_trial, delta, tree_ref}. `status` (kept|reverted|rebased|crash) is what
+        drives the attempt-loop branch in protocol.md "Isolated attempt loop"; by the
+        time you read it the tree action has already happened. Never branch on the process
+        exit code: it carries the T0 binary (0 green / 1 red / 2 bad argv, the oracles/*
+        convention), so a `kept` red-but-improved attempt, the ratchet's own signature case,
+        exits 1. On red, its `discovered_tasks` field is
+        the AEGIS digest to fold into the next brief — no separate digester dispatch needed
+        (harness verify t0 calls its sibling harness probe digest internally on failure).
+```
+
+## 4. EVAL → spec-evaluator (once per round)
+```
+compile-order --operation evaluate --slug <slug> --worker spec-evaluator --round <r>
+  --payload '{"dimensions": ["spec-conformance"], "run_cmd": "<cmd>", "t0_artifacts": [...]}'
+Invoke via Agent (model: eval), ONCE, after GATE L2:
+  Skill(shapeup-sdlc-plugin:spec-evaluator) --order <path>
+Effect: one feature-level pass over the running app against all AC + Done-when; writes
+        evaluation/EVAL-FEATURE-<slug>.md (verdict + bug list) + its WorkResult (criteria
+        verdicts, refuted boxes, T0 citations). It touches NO task file and NO board.
+ingest-result <results/evaluate-r<r>.json>: appends the .verdicts JSONL ledger, un-ticks the
+        refuted AC boxes, sets eval_verdict frontmatter — the judge returns data, ingest writes.
+Read back: EVAL-FEATURE-<slug>.md → verdict (pass|fail) + the bug list (each bug has
+        task ref, severity, file:line, expected vs actual).
+```
+
+> Dependency note: this uses spec-evaluator's **feature-level** pass (`--feature <slug>`),
+> which evaluates the whole board in one session rather than one task at a time. If your
+> installed spec-evaluator is the per-task v0.1, add the `--feature` mode (a small v0.2
+> patch: iterate the board's AC/Done-when in one probe+grade session, emit one
+> EVAL-FEATURE report) before wiring the tech lead to it. The per-task invocation still
+> works for ad-hoc checks, but the round loop expects one feature pass.
+
+## 5. SHIP / GATE H → scope-hammer
+```
+Invoke via Agent (model: exec): Skill(shapeup-sdlc-plugin:scope-hammer)
+        --slug <slug> [--baseline <path>] [--breaker outer|inner --scope <id>]
+Effect: GATE H0 census (scopes + QA findings + discovered ledger + advisor-overflow flags) →
+        H1 baseline comparison (never vs. a perfect ideal) → H2 cut list + verdict.
+Read back: the proposed cut list + verdict (SHIP now | SHIP after fixing ship-blocking items |
+        CANNOT SHIP). The tech lead records the PO's decision in round-ledger.md and performs
+        the actual close (SHIP S.1 onward) — scope-hammer proposes, it never ships.
+```
+
+## Authority boundaries (do not cross)
+- The Scout orients; it never plans, builds, or judges — it hands raw material to the planner.
+- The planner decides scope; the tech lead confirms it with the PO at GATE L1b.
+- The generator reports task outcomes in its WorkResult; `harness reduce ingest` flips `status:
+  done` from that report. The tech lead confirms the feature-level close at GATE L4.
+- The evaluator issues verdicts only; it never closes tasks. Judge ≠ doer.
+- The tech lead decides *when* and *whether* each skill runs and how rounds proceed, owns
+  run-state + the Hill report — it does not decide *what* a task contains or *whether* a
+  single AC passes.
+
+## Handoff files (the shared state)
+| File | Written by | Read by |
+|------|-----------|---------|
+| `<name>.en.md` + `glossary.md` | translator (L0, if non-English) | tech lead (ORIENT input), Scout, planner |
+| `orient/code-surface.md` | Scout (step 7) | planner Phase 1 (ingest, no re-scan) |
+| `orient/spike-<area>.md` | Scout (step 7) | planner Phase 1b/contracts; tech lead (L1a) |
+| `orient/discovered-seed.md` | Scout (step 7) | planner Phase 6 (task gen) |
+| `orient/hill-signal.md` | Scout (step 7) | tech lead (renders L1a Hill) |
+| `orders/<id>.json` (WorkOrder) | **harness compile (mechanical)** | every worker (its ONLY pipeline input), validate-envelope hook |
+| `results/<id>.json` (WorkResult) | the dispatched worker (its ONLY pipeline output) | harness reduce ingest (the single writer of everything below it) |
+| `.shapeup/<slug>/tasks/*` (LOCAL board, v3.2) | **harness reduce ingest** (status, AC ticks, unblocks) + planner orders (task bodies) | tech lead (board status), compile-order (next task, ACs) |
+| `discovery/ledger.md` | **harness reduce ingest** (from workers' `discoveries[]`) | reconcile orders (ba), scope-hammer (H0 census) |
+| `scope-summary.md` | planner (analyze/reconcile orders) | tech lead (Done-when), evaluator (Done-when criteria) |
+| `evaluation/EVAL-FEATURE-<slug>.md` + `.verdicts-*.jsonl` | evaluator (report) / **harness reduce ingest** (verdict ledger, un-ticks) | tech lead (verdict), next fix order (bug list) |
+| `harness-run.md` | **tech lead (sole writer)** | tech lead (round ledger + Hill + run-state), PO (audit) |
+| `scopes/<scope-id>.md` | `scope-architect` (sole writer) | tech lead (substrate/sequence), sandbox hook (write-whitelist), compile-order (inlined into orders) |
+| `t0/verdicts/r<N>-a<M>-t<T>.json` | `harness verify t0` (skill-local, mechanical — not a worker) | spec-evaluator (required citation), tech lead (hill derivation), compile-order (digested errors) |
+| `t0/trials.jsonl` (the ratchet ledger, append-only, `baseline_trial` as the parent link) | `harness verify t0` (one row per attempt: score, status, delta, tree_ref) | compile-order (`trial_history` into the next order), ship-report (T0 + Ratchet sections), `harness probe stats --ratchet` |
+| `round-ledger.md` | **tech lead (sole writer)** | compile-order (decisions into every order), PO (audit) |
+| `hill/<scope-id>.yml` + `hill-chart.md` | **tech lead (sole writer)** | PO ("status without asking"), scope-hammer (H0 census) |
+
+> The single-writer rule is closed mechanically: no worker writes `run-state.md`, the board, or the ledger. A worker
+> performing a shared-state write is a defect — route it back through its WorkResult.
+
+---
+
+# Part 3 — The invariants
+
+Moved out of `SKILL.md` when the round loop became a launchable script:
+most of what this table used to guard against — a partial board reaching EVAL, a gate crossed
+on the model's own authority, an evaluator called mid-BUILD — is now a property of
+`skills/tech-lead/workflows/shapeup-run.js`'s code, not a rule a model has to remember to obey.
+The table stays as the readable rationale for WHY the code is shaped the way it is; the runtime
+guarantee lives in the script and, where noted, in a hook.
+
+| Rule | Rationale |
+|------|-----------|
+| Orchestrates Building only (steps 7–11); shaping/betting/kick-off are PO-personal, upstream | Intake is a kicked-off pitch, not a raw idea — the tech lead does no shaping/planning-authority work |
+| ORIENT (step 7) runs before MAP SCOPES (step 8) | Roadmap: no pre-divided tasks at kick-off; the team orients first so the board is reality-born |
+| Intake must be English before ORIENT; tech lead does NOT translate — it delegates to `translator` at GATE L0 | Translation is a separate single-purpose skill; the intake conversation only detects + sequences it, before RunArgs is ever compiled |
+| Every worker dispatch goes through the envelope port: compile-order → `--order` → ingest-result; shared state is written ONLY by ingest | The single-writer rule is mechanically true: a worker that writes boards/ledgers/run-state is a defect, and a malformed envelope is denied by the validate-envelope hook before it can corrupt run truth. `shapeup-run.js` uses this same shape for every operation in the central registry — orient, wire, analyze, map-scopes, evaluate, hunt, hammer alike |
+| Progress is reported by Hill position, never by counting tasks | The roadmap forbids task-counting; a 90%-done slice can still be stuck uphill on the one unknown that matters |
+| Evaluator runs once per round, only after GATE L2 (board 100% done) | The whole point: cheap end-of-round QA, never per task. `shapeup-run.js`'s round loop dispatches spec-evaluator exactly once per iteration, after the GATE L2 resolution — there is no code path that calls it from inside the scope attempt loop |
+| Evaluator never called inside the BUILD loop | Keeps the build coherent and the run cheap |
+| r>1 builds bugs only, never the whole board | Don't re-do passing work; minimize churn — see protocol.md's regression rule for what DOES re-run (touched UCs' full Test Surface) |
+| Stop at max_rounds; escalate honestly | No infinite fix loops; `shapeup-run.js` returns `{status: "gate_h", breaker: "outer"}` rather than looping past the budget |
+| Tech lead delegates, never reimplements a sub-skill | Stays thin; each skill keeps its own gates and authority |
+| Every delegation to a sub-skill (except the mechanical `harness verify t0`/`harness compile`/`harness reduce ingest`) is a fresh Agent on the L0.8-resolved model | Isolation the zero-memory-handoff design assumes; a direct inline call would silently drop the model matrix — see references/protocol.md "Invocation mechanism" |
+| Planner stays high-level on tech | Spec errors cascade into every build round |
+| Never auto-deploy; "shipped" never silently means "deployed" | Deploy is outward-facing, PO-gated; record "deploy pending (PO)" otherwise |
+| "Shipped" names the dims NOT evaluated | `RunReturn`'s `dims_not_evaluated` field carries this; the L4 sign-off block shows it, never silently drops it |
+| Every gate emits the canonical `⏸ GATE LN — Title` block before any narrative | Composed by the workflow (`gateBlock()`), emitted VERBATIM by the skill — conversational re-summary is not a gate |
+| In interactive/--auto: a `paused` return stops and waits for PO confirmation | Never auto-proceed past a gate; the PO must cross each threshold explicitly — see "The pause protocol" in SKILL.md |
+| At GATE L3 FAIL: name scope (task + failed criterion), never prescribe fix options | Root cause analysis and fix paths belong to the implementer, not the orchestrator |
+| SHIP harvest records facts only — copies existing structured output, never computes a new verdict/score | A self-computed score = a second judge behind spec-evaluator (breaks single-judge, invites Goodhart); the eval suite interprets, harvest records |
+| Three-level circuit breaker: attempt_budget (inner, per scope) nests inside round_budget (outer), with an opt-in wall_clock_budget_s deadline | An exhausted scope queues a GATE H hammer proposal, it never blocks the round; only round_budget hitting 0 stops the whole run; the deadline breaker (checked every round boundary in `shapeup-run.js`) routes to GATE H so a run out of clock still ships what is green instead of being killed from outside |
+| The tech lead never hand-edits a scope contract | scope-architect is its sole writer (single-writer-per-file) |
+| Substrate-disjointness + PA1/PA2 lints are re-asserted at GATE L1b (harness verify spec) even when scope-architect already checked them | A human may have hand-approved past a 🔴 at the architect's checkpoint; `shapeup-run.js` runs spec-lint itself, in code, before resolving L1b |
+| Hill phase is read from mechanical facts (T0/T1/seesaw), never declared by a worker | Closes the self-reported-confidence risk outright — facts move dots, not authors |
+| GATE H is delegated to scope-hammer, never adjudicated inline by the tech lead | Keeps the orchestrator thin; census/baseline-comparison/cut-list logic has one owner |
+
+---
+
+# Part 4 — State
+
+Two halves of one question, folded into this file when the references consolidation retired the
+standalone `state.md`: who owns which artifact (state ownership), and the shape of what the tech
+lead itself writes — `harness-run.md`, the run's own ledger.
+
+## State ownership (D6, mechanically closed in v1.0)
+
+Workers are stateless; the orchestrator layer is the **sole writer of ALL run-state**. Every
+worker receives a structured **WorkOrder** envelope (`.shapeup/<slug>/orders/`, compiled by
+`harness compile`) and returns a **WorkResult** envelope (`results/`); the deterministic
+`harness reduce ingest` performs every shared-state write — board status, AC ticks, unblock
+propagation, discovery-ledger appends, verdict bookkeeping.
+
+No worker writes `run-state.md`, `tasks/_index.md`, or the ledger. Everything a worker used to
+write into shared files, it now returns as data.
+
+The tech lead owns `harness-run.md` — rounds, gate decisions, Hill positions, verdicts,
+`discovered_rounds`, config, language record. The board (`tasks/_index.md`, LOCAL root — v3.2) is
+**execution truth**, maintained exclusively through ingest.
+
+**The run receipt (v1.4).** `harness init run` opens the run and writes
+`.shapeup/<slug>/receipt.json` plus `.shapeup/active-scope` before any gate. The
+receipt is the mechanical fact that a run *started* — distinct from every other artifact here,
+which records what a run *did*. That distinction is load-bearing: the guards that check a run's
+progress (the ship report's census) are all scoped to an active run, so before the receipt
+existed, a run that never started was invisible to every one of them. `hooks/gate-zerowork.mjs`
+reads only the receipt's presence, which is why it can see a total failure that leaves no other
+trace. See `references/gates.md` — GATE L0.1.
+
+## Central domain registry
+
+Every record type and payload field that crosses a skill boundary is defined exactly once in
+`skills/tech-lead/schemas/domain.schema.json` — the envelope schemas (`work-order.schema.json`,
+`work-result.schema.json`) only `$ref` it. The registry annotates each entity's tier
+(SHARED/LOCAL), location, sole writer, and readers, carries the machine-readable ERD (`x-erd`),
+and maps which payload fields each worker may rely on (`x-payload-by-worker`).
+
+A new cross-boundary field is added THERE first (structural test #24 enforces the map's
+consistency); a skill inventing its own undeclared field is a defect — the orchestrator, not the
+worker, owns the vocabulary.
+
+## Two ledgers, split by promotion timing
+
+(Addendum §F.3 — only when scope contracts exist.)
+
+`harness-run.md` stays the LOCAL (`.shapeup/<slug>/`, gitignored) full run trace: it can be
+rebuilt or lost without consequence.
+
+A second, committed `round-ledger.md` (`shapeup/<slug>/round-ledger.md`, SHARED,
+Tier A) holds only what must survive a crash or a `.shapeup/` wipe:
+
+- the resolved model/budget matrix (L0.8/L0.9),
+- the **Decisions** table — every gate crossing and every PO answer to a blocked phase,
+  promoted the instant it is given, never batched to round close.
+
+Gate crossings resolved from a **gate answer set** (`harness gate`) are written here
+with their source — `preset:ci`, `file:.shapeup/gate-answers.json` — and the set's
+`authorized_by`. A headless run that ships must always be able to name the human behind its
+sign-off; that name lives here and nowhere else.
+
+The tech lead is the sole writer of both. `round-ledger.md` is simply the subset that must never
+live only in a session or a gitignored file. No scope contracts → `round-ledger.md` is not
+written, and `harness-run.md`'s "Decisions log" is the only ledger, exactly as in v0.2.6.
+
+## The ledger schema
+
+The tech lead's run record. One per feature, lives in the LOCAL run-trace root
+`.shapeup/<slug>/harness-run.md` (hidden, gitignorable — it is ephemeral run-state,
+not a shared deliverable). It is the structured
+artifact that carries state across rounds and across sessions (so `--from` can resume), and
+the PO's audit of how the feature was built.
+
+**The tech lead is the sole writer of this file.** It is the authoritative
+run-state for the whole build phase — rounds, gate decisions, Hill positions, verdicts,
+`discovered_rounds`, config, language record. Workers never write here; the tech lead passes
+them what they need (`feature`, `spec`, `stack`, `discovered_rounds`) as args. The board
+(`tasks/_index.md`) remains the planner/generator's execution truth that the tech lead reads.
+
+## Frontmatter
+```yaml
+---
+type: harness-run
+feature: [slug]
+spec_folder: [path to SHARED spec deliverable, e.g. shapeup/<slug>/spec/]
+lens: lite | standard | cross-context
+eval_dimensions: [spec-conformance]   # the set from GATE L0.5 (init-run --dimensions); every EVAL order is compiled from THIS line
+max_rounds: 3
+auto_level: interactive | auto | unattended
+status: orienting | mapping | building | evaluating | shipped | escalated
+final_verdict: ~ | pass | fail | not-evaluated
+rounds_used: [N]
+discovered_rounds: [N]
+deploy: ~ | deployed | pending-po
+started_at: [ISO]
+closed_at: ~ | [ISO]
+---
+```
+
+## Round table (the spine)
+Mirrors the long-running harness cost table — one row per phase, so you can see where time
+and tokens go and that EVAL is cheap relative to BUILD.
+
+```
+## Rounds
+| Phase            | Round | Result            | Duration | Notes |
+|------------------|-------|-------------------|----------|-------|
+| Orient           | —     | spiked VNPay seam | 18 min   | spike resolved; 9 discovered tasks seeded |
+| Map Scopes       | —     | 14 tasks, 5 layers| 5 min    | orient-informed; 1 SPIKE (VNPay) |
+| Build            | 1     | 14/14 ✅          | 2 h 10 m | all tasks closed |
+| Eval             | 1     | FAIL — 3 bugs     | 9 min    | EVAL-FEATURE-checkout-vnpay.md |
+| Build            | 2     | 3 bugs fixed      | 22 min   | bug-only re-build |
+| Eval             | 2     | PASS              | 8 min    | verdict pass |
+| Ship             | —     | built & verified  | —        | dims: spec-conformance; deploy pending (PO) |
+```
+
+## Hill report (the progress narrative — NOT task counts)
+The roadmap forbids reporting progress by counting tasks: a 90%-done slice can still be stuck
+uphill on the one unknown that matters. So the tech lead reports each slice's **position on
+the hill**, derived mechanically from open unknowns. Render it at every round boundary
+(area-level at GATE L1a — before slices exist; slice-level from GATE L1b onward).
+
+```
+## Hill — round 1
+| Slice        | Position    | Derived from |
+|--------------|-------------|--------------|
+| S1-spine     | 🔽 downhill | spike closed + spine render-AC passed (crest crossed); 2 known tasks left |
+| S2-filters   | 🔼 uphill   | SPIKE-003 open: pagination approach unproven |
+| S3-export    | 🔼 uphill   | contract ⏳ TBD: file-format field |
+```
+
+Position triggers:
+- 🔼 **Uphill** — open SPIKE / `⏳ TBD` contract / unresolved discovered task / approach unproven.
+- ⛰️ **Crest** — all unknowns resolved AND a concrete board fact (the spine slice's render-AC passes).
+- 🔽 **Downhill** — only known work remains, no open unknowns.
+- ✅ **Done** — slice clickable-done.
+
+Source: `orient/hill-signal.md` (area-level, at L1a) then the board + open SPIKE/contract state
+(slice-level, L1b onward). If slice IDs aren't on the board yet, report at
+task-group level and note the fallback here.
+
+## Decisions log
+Every L-gate decision, for traceability.
+```
+## Decisions
+- GATE L1a: Orient accepted; spiked VNPay seam (resolved); S2-filters flagged uphill.
+- GATE L1b: PO accepted board; cut "guest checkout" task to phase 2.
+- GATE L2 (r1): board green; ran eval (feature not trivial).
+- GATE L3 (r1): FAIL → approved bug-only round 2.
+- GATE L3 (r2): PASS → ship.
+- GATE L4: built & verified; deploy pending (PO); closed.
+```
+
+## Bug carry-over (when FAIL)
+Links the current round's bug list so the next BUILD round has its worklist.
+```
+## Open bugs (round 1 → fix in round 2)
+→ see evaluation/EVAL-FEATURE-checkout-vnpay.md
+- BUG-1 critical SC-DONE-WHEN apps/web/checkout/Pay.tsx:84
+- BUG-2 major   SC-ERR     apps/api/payments/handler.ts:51
+- BUG-3 minor   SC-NONGO   apps/api/shared/auth.ts:12
+```
+
+## Escalation block (only if max_rounds hit without PASS)
+```
+## Escalation
+Rounds used: 3/3 — still FAIL. Residual bugs: [N].
+Recommendation: [cut scope of feature X | accept minor bugs | extend max_rounds with PO approval].
+Decision owner: PO.
+```
+
+## Rules
+- The ledger is append-mostly: each round adds rows, never rewrites history.
+- `status` + `final_verdict` are the resume anchors for `--from`.
+- A `not-evaluated` final verdict (from `--no-eval`) is recorded plainly — never silently
+  upgraded to `pass`.
+
+## `round-ledger.md` (committed, Tier A — scope contracts only)
+
+Lives at `shapeup/<slug>/round-ledger.md` (SHARED root, tracked). Not a replacement
+for `harness-run.md` — a small, committed **subset** of it: the two things that must survive
+a `.shapeup/` wipe or a crash. Absent on specs with no scope
+contracts; `harness-run.md`'s existing Decisions log stays the only ledger there.
+
+```yaml
+---
+type: round-ledger
+feature: [slug]
+models:                          # L0.8 resolved matrix, recorded once, source noted
+  orch: [model] (source: flags|settings.local|settings.json|default)
+  exec: [model] (source: ...)
+  eval: [model] (source: ...)
+  qa:   [model] (source: ...)
+  digester: script | sonnet
+budgets:
+  round_budget: [N]              # outer breaker
+  attempt_budget: [N]            # inner breaker, per scope
+---
+```
+
+```
+## Decisions
+| Round | Scope | Kind | Question | Answer | Resolved by |
+|-------|-------|------|----------|--------|-------------|
+| 2 | cart-creation | design-decision | cart-total rounding | round half-up | PO (interactive) |
+| 2 | cart-creation | substrate-expansion | needs packages/shared/http.ts | approved → shared_substrate | PO (interactive) |
+```
+**Promotion timing:** a row is appended the INSTANT the PO answers —
+never batched to round close. This is the file `task-executor`'s isolated briefs read back
+(the zero-memory handoff): an answer given once in round 2 must still be known in round 5's
+fresh-context attempt without replaying any chat history.
+
+## Harvest row — `metrics/<machine-id>.jsonl` (written at SHIP)
+
+`harness-run.md` is ephemeral run-state — needed live for `--from` resume, the
+FAIL-loop, and QA reconcile; worthless after ship. At SHIP the tech-lead **harvests**
+the durable-mineable *signals* out of it into one append-only row:
+
+```
+.shapeup/metrics/<machine-id>.jsonl   # one row = one e2e run; LOCAL (gitignored)
+```
+
+Path note: `metrics/` lives under the LOCAL root with the rest of the run trace — moved
+there by ADR-0001, because a committed shard keyed on `$HOSTNAME` only grows and puts a
+person's machine name in the repo. It outlives any single run (the per-slug run dirs are
+superseded run by run; the shard only appends), which is what makes it the durable signal
+feed. Sharded per machine so shards can be pooled deliberately without colliding on one
+filename; the read plane is `harness probe stats`, or `cat .shapeup/metrics/*.jsonl`.
+`schema_version` makes an old row readable by later skill versions.
+
+### Two hard rules (same discipline as the Test Surface: *derived, never invented*)
+1. **Harvest only fields that already exist as structured output at ship time.** If a
+   field forces the tech-lead to *evaluate something new* → reject (judgment in disguise).
+2. **Harvest records facts, never computes a new verdict.** A self-computed
+   `run_quality_score` would be a second judge behind `spec-evaluator` → breaks the
+   single-judge rule and invites Goodhart. The eval suite *interprets* downstream;
+   harvest *records*.
+
+Scope: the harvest feeds **only the e2e pipeline measurement**. The per-skill functional
+fixtures run in isolation and do not consume it.
+
+### Row schema (one JSON object per line)
+| Field | Existing source (copied, never re-graded) | Signal |
+|---|---|---|
+| `schema_version` | constant `1` | forward-compat |
+| `run_id` | `receipt.json` (copied verbatim) | **the join key** — the row's only link to the run trace that produced it, and to the S.7 export. Every other field here is a count, and `feature_slug` groups two runs of one feature together rather than apart |
+| `feature_slug` | run-state frontmatter | identity |
+| `terminal_state` | run-state final: `shipped` / `circuit_broken` / `abandoned` | circuit-breaker outcome |
+| `round_count` | round table | effort-to-PASS |
+| `final_audit_score` | final EVAL report (copied, not re-graded) | conformance |
+| `surprise_count` | `.shapeup/<slug>/discovery/ledger.md` | shaping quality — scope drift |
+| `spike_unresolved_count` | `SPIKE-UNRESOLVED` markers at bet | shaping quality — open risk into bet |
+| `scope_cut_count` | `~` items cut at SHIP S.0 | appetite pressure / scope hammer |
+| `qa_findings` | `.shapeup/<slug>/qa/hunt-report.md` + triage → `{total, promoted, held}` | edge quality |
+| `slice_count` | breadboard B5 (≤9) | **normalizer / denominator** |
+| `sources` | path to each **SHARED** source artifact — never a LOCAL `.shapeup/` path (the run-trace is superseded run by run, so a LOCAL path dangles by the time anyone reads the row; SHARED paths resolve on any clone — tier-direction rule) | auditability |
+
+- `slice_count` is the **denominator**: `round_count=4` on a 2-slice feature is alarming,
+  on a 9-slice feature is normal. Without it, e2e comparisons are apples-to-oranges.
+  Enables `round-per-slice`, `surprise-per-slice`.
+- `spike_unresolved_count` + `surprise_count` measure two of the three downhill
+  conditions (open-risk-remaining, scope-drift-from-breadboard). A good shaping run
+  drives both toward 0 — measured from the build trace, no manual grading.
+- **Rejected fields:** `time_spent` / velocity (no clock; Shape Up forbids counting hours
+  — `round_count` is the legitimate effort proxy) and `run_quality_score` (second judge).
+  Even a real cost or wall-clock record would not belong here: a signal feed that carried a
+  duration would become a velocity feed on the next person who read it. This harness does not
+  currently carry one — the design called for one derived from a per-agent-call journal, but
+  nothing in the pipeline ever writes it, so there is no such record to reject or admit.
+
+### Row template
+```json
+{"schema_version":1,"run_id":"checkout-vnpay-20260813T091233Z-c714ea8d","feature_slug":"checkout-vnpay","terminal_state":"shipped","round_count":2,"final_audit_score":"PASS","surprise_count":3,"spike_unresolved_count":0,"scope_cut_count":1,"qa_findings":{"total":5,"promoted":1,"held":4},"slice_count":4,"sources":["shapeup/checkout-vnpay/shaping/shaping.md","shapeup/checkout-vnpay/shaping/breadboard.md"]}
+```
+
+LOCAL artifacts (the EVAL report, discovery ledger, QA hunt report) are *harvest-time reads*:
+their **values** are copied into the row's fields (`final_audit_score`, `surprise_count`,
+`qa_findings`) but their paths are never recorded in `sources` — they are gitignored and
+wiped, so a committed pointer to them is dead on arrival.