@azure-id/orc 0.55.0

package/mock-run/a-normal-day.md

@@ -0,0 +1,587 @@
+# A normal day — what it looks like when all six new lanes are installed
+
+> One normal day. One normal run. `/orc-pact`, `/orc-boundary`, `/orc-handoff`,
+> `/orc-budget`, `/orc-aftermath` and `/orc-export` are all installed, plus
+> the wiki partial refresh.
+
+This is the file that answers: **so what actually changes?**
+
+The other six files each show one lane alone. That is not how you will meet
+them. You will meet them all at once, inside a run you were going to do anyway.
+
+---
+
+## 1. The setup
+
+```
+$ orc --version
+orc 0.53.0
+
+$ orc doctor
+  ✓ install manifest current   (40 agents, 29 skills)
+  ✓ settings wiring present
+  ✓ trace pointer clean
+  ✓ AGENTS.md current
+```
+
+Same project as the other files: **shopcart**, a small web shop.
+Same two people: **Rina** (developer) and **Budi** (product manager, does not
+read code).
+
+Config, all defaults except two:
+
+```yaml
+boundary_gate: block          # Rina moved this up from `warn` after one month
+wiki_scan_tier: ladder
+```
+
+Today's job: **"let customers take a refund as store credit instead of money
+back."**
+
+This touches the front end, the back end, the database, and money. It is a
+good test, because it hits every new lane at once.
+
+---
+
+## 2. The run, from start to end
+
+### Phase 1 — preflight
+
+This is the screen that changed the most. Six lines instead of three.
+
+```
+> /orc  add store credit as a refund option
+```
+
+```
+ORC · run-orc-store-credit-081026-093012
+
+Preflight
+  wiki       : FRESH — 14 docs · 8 in active use · 2 pending refresh, est. $0.14
+  pattern    : typescript cached (02-09-2026)
+  crosslink  : none configured
+  pact       : 5 holding · 1 drifted · 0 uncheckable          ← 01
+  boundary   : 3 areas execute · 1 escalate · 1 refuse        ← 02
+  aftermath  : api/refunds/ churned after the last run here   ← 05
+
+Three of those need a word before we start.
+
+  1  PACT-004 is DRIFTED.
+     "A refund is only allowed inside 45 days."
+     api/refunds/window.ts changed 4 days ago and we owe it a re-check.
+     It takes 4 seconds and costs nothing. I will run it now unless you say no.
+
+  2  db/migrations is REFUSE.
+     Store credit almost certainly needs a table. I will plan the migration
+     as a draft that nobody runs, and hand it to you at the end.
+     (This is boundary_gate: block. Set it to warn to change this.)
+
+  3  api/refunds/ churned after run-orc-tax-rounding-190826 — the file was
+     rewritten twice in the 8 days after we shipped there. Something about
+     that area is harder than it looks. I will tell the analyst.
+
+  Continue?  [yes] / stop / change something
+```
+
+```
+> yes
+```
+
+```
+Preflight · re-check
+
+  PACT-004  npm test -- refund-window     ✓ pass   → HOLDING
+            re-anchored at 3b71ce2
+
+  pact now: 6 holding · 0 drifted · 0 uncheckable
+```
+
+**Note what happened.** A free 4-second check cleared a stale promise before
+one token was spent on planning. Last year that promise would have sat
+forgotten and the planner would never have known about it.
+
+**And PACT-006 is gone from the uncheckable list.** In the `/orc-aftermath` mock it had been
+uncheckable for 47 days, flagged by three lanes. Someone finally wrote the
+test. This is what the flywheel looks like when it works.
+
+---
+
+### Phase 2 — intake and the price tag
+
+```
+ORC · intake
+
+The analyst asked 4 questions. You answered. Spec is settled.
+The planner produced 14 tasks in 5 waves.
+
+Two promises point at files this plan touches. They are now hard constraints
+for the planner:                                                        ← 01
+
+  PACT-004  refund window is 45 days (temporary, review 01-12-2026)
+  PACT-006  refund money goes back to the original payment method
+
+  ⚠ PACT-006 is interesting here. "Store credit" is NOT the original payment
+    method. This plan may break a promise you made in July.
+
+    Is store credit an exception to PACT-006, or does PACT-006 need to change?
+
+      1  It is an exception. Store credit is opt-in, the customer chooses it.
+      2  PACT-006 needs to change. Rewrite it.
+      3  Stop and think about this.
+```
+
+```
+> 1
+```
+
+```
+  Recorded. PACT-006 gains an exception clause and a new test case is added
+  to the plan (T13) that proves the exception is explicit and opt-in.
+  A promise was not quietly broken. That is the whole point of the ledger.
+```
+
+Then the budget:
+
+```
+ORC · intake · cost                                                     ← 04
+
+  14 tasks · 5 waves · pause after every 2 waves
+  Your plan: Max 20x. Primary unit: quota.
+
+  lane          raw tok   weighted   quota    usd p50   time      note
+  ──────────────────────────────────────────────────────────────────────────
+  /orc-ultra     3.10M     1.38M      45%     $16.80    38–62m   more than needed
+  /orc           1.64M      752k      24%      $7.90    22–36m ← recommended
+  /orc-mini      0.58M      266k       9%      $2.60    11–18m   no review on money
+  /orc-fast      not possible                                    1 REFUSE, 2 ESCALATE
+
+  T12 (the migration) is REFUSE and will not be dispatched.
+  Without it: 1.34M raw · 614k weighted · 20% of window · $6.45 → $10.10.
+
+  token split, p50, on /orc
+    input        71k   ( 4%)      cache write   484k  (30%)
+    cache read  979k   (60%)      output        103k  ( 6%)
+    cache read is 60% of the count and 12% of the cost. Budget on the
+    weighted number, not the raw one.
+
+  Confidence: good in 6 bands, fair in 1, LOW in 1.
+  Band [90,100] has 4 samples. The top of the range is soft.
+  Unattributed in the corpus: 2.4%.
+
+  Context: no task above 78% of its window. Nothing at risk.
+
+  1  Start on /orc      614k weighted · 20% of window · $6.45 → $10.10
+  2  Switch to /orc-mini
+  3  Change the plan first
+```
+
+```
+> 1
+```
+
+**Before all this, that screen said:** *"14 tasks, 5 waves, pause after 2.
+Continue?"* Now it has a token ledger, the unit Rina is actually billed in, a
+lane comparison, a context-risk check, and an honest note about which numbers
+are soft.
+
+---
+
+### Phase 3 — waves
+
+```
+Wave 1 of 5 — 3 tasks
+
+  T01  TDD: red proof for store-credit refund      sonnet-5-high   [EXECUTE]
+  T02  TDD: red proof for the PACT-006 exception   sonnet-5-high   [EXECUTE]
+  T03  Add the store_credit type to the API        sonnet-4-6-high [EXECUTE]
+
+  wiki: orc-feature-refunds.md → T01, T02, T03
+```
+
+```
+Wave 3 of 5 — 3 tasks
+
+  T08  Call the credit ledger on refund      opus-4-8-high   [ESCALATE]
+       card: api/refunds — this touches money and PACT-006's new exception.
+       Rina must read this before ship.
+
+  T09  Show the store-credit option in the UI   sonnet-4-6-high [EXECUTE]
+
+  T12  Create the store_credit table            NOT DISPATCHED  [REFUSE]     ← 02
+       db/migrations card: no down-migration, no staging database.
+       Lifted out of the wave. The wave continues with 2 tasks.
+
+       I will still write db/migrations/0011_store_credit.sql as a DRAFT.
+       Nobody runs it. It is yours to review and run.
+```
+
+At the pause:
+
+```
+Pause after wave 2 of 5
+
+  Burned so far : 267k weighted · 8% of window · $2.14                    ← 04
+                  (forecast to here 238k–390k · 7–12%)
+  Left to go    : 347k → 610k weighted · 11–19% of window
+  Inside the forecast.
+
+  Waves done    : 2 of 5
+  Tasks done    : 6 of 13 dispatched (1 refused)
+
+  1  Continue     2  Stop here     3  Show me what changed so far
+```
+
+---
+
+### Phase 5 and 6 — review, verify, promises
+
+```
+Phase 5 · review        12 files · 3 findings (1 P1, 2 P3)
+                        P1: the store-credit path skips the idempotency guard
+                        → sent back to T08, repaired, re-reviewed, clean
+
+Phase 6 · verify
+  build                 ✓
+  tests                 ✓  158 passed (14 new)
+  TDD gate              ✓  both red proofs went green
+  acceptance criteria   ✓  6 of 6
+
+  pact re-check — only what this change touched                          ← 01
+    PACT-001  payments never double-written     ✓ HOLDING
+    PACT-004  refund window 45 days             ✓ HOLDING
+    PACT-006  refund to original method         ✓ HOLDING
+              exception recorded: store credit, opt-in, proved by T02
+
+  boundary re-check                                                      ← 02
+    api/refunds  ESCALATE → still ESCALATE. Rina must read T08.
+    db/migrations REFUSE  → unchanged. The draft migration is waiting.
+```
+
+---
+
+### Phase 7 — ship, and the five new closing lines
+
+```
+ORC · ship
+
+  Shipped 12 tasks. 19 files changed. Branch: feat/store-credit
+
+  ── Before you merge ────────────────────────────────────────────────
+
+  1  Read this one yourself                                            ← 02
+     T08  api/refunds/credit.ts — the credit ledger call.
+     Why: it touches money and it is the first code under PACT-006's
+     new exception. The reviewer passed it. A human should still look.
+
+  2  Handed back to you, not attempted                                 ← 02
+     T12  Create the store_credit table.
+     Draft written: db/migrations/0011_store_credit.sql — NOT run.
+     To let ORC do this next time:
+       ☐ add 0011_store_credit.down.sql
+       ☐ add staging.database_url to config/app.yaml
+       ☐ add a CI job that runs migrations up and down
+     Then `orc boundary status db/migrations` will say ESCALATE.
+
+  ── Small things ────────────────────────────────────────────────────
+
+  3  2 of these files were GREEN self-serve surfaces                   ← 03
+       web/locales/en.json
+       web/locales/id.json
+     Next time a product manager can change those text strings alone,
+     with /orc-handoff or the Self-serve panel. No run needed.
+
+  4  The wiki is now 3 docs behind — 79k tokens / $0.20                ← wiki
+     This run changed api/refunds/ and web/checkout/.
+       orc-feature-refunds.md   TOUCHED  6 files  used 17/20  deep   46k  $0.14
+       orc-feature-checkout.md  TOUCHED  2 files  used  9/20  light  21k  $0.04
+       orc-reference-config.md  TOUCHED  1 file   used  2/20  light  12k  $0.02
+     Free first:  [ orc wiki sync ]  — registration only, 0 tokens
+     Then:        /orc-wiki refresh --top 1   (46k covers 17 of 28 recorded uses)
+
+  5  AGENTS.md is stale — 2 sources moved                              ← 06
+     Cursor and Codex on this team are reading the old version.
+     Fix: orc export     (3 seconds, no tokens)
+
+  ── Cost ────────────────────────────────────────────────────────────
+
+              forecast p50          actual          diff                 ← 04
+  raw          1.34M tokens         1.47M           +10%
+  weighted       614k                670k           +9%   inside p90
+  quota           20% of window       22%
+  usd           $6.45               $7.02
+  time          22–36 min           27 min
+
+  cache read    60% of raw forecast → 58% actual   (cache behaved)
+  output        103k forecast → 129k actual  (+25%) ← the repair round
+
+  Band [80,90) ran +22% weighted — T08 needed one repair after the P1
+  finding, and a repair is almost pure output tokens. That is why the
+  money moved more than the token count did.
+
+  Fed to /orc-retro. That band now has 7 samples.
+```
+
+---
+
+## 3. What changed, line by line
+
+Here is the same run, before and after, side by side.
+
+| Moment | v0.45.0 (today) | v0.53.0 (all shipped) |
+|---|---|---|
+| Preflight | 3 lines: wiki, pattern, crosslink | 6 lines. A stale promise is re-checked free. A REFUSE area is named. A past churn warning is passed to the analyst. |
+| Intake | "14 tasks, 5 waves. Continue?" | Plus a token ledger per lane in the unit you are billed in, a context-risk check, a confidence note, and two promises turned into planner constraints |
+| **A promise about to break** | Nobody notices. It breaks in November. | ORC asks one question at Phase 2 and adds a test |
+| Wave dispatch | All 14 tasks dispatched | 13 dispatched. The one nobody could safely run is lifted out and drafted instead |
+| Pause | "Continue?" | Plus spend so far vs forecast |
+| Verify | build + tests + criteria | Plus a re-check of only the promises this change touched |
+| Ship | "Shipped 12 tasks" | Plus what to read, what was handed back with a checklist, what a PM could have done alone, the wiki debt with a price, and a stale export warning |
+| After ship | nothing | Cost forecast vs actual, fed to retro. Aftermath will grade this run in 7 days. |
+
+---
+
+## 4. The rhythm — a month with all of it
+
+None of these lanes are things you run every day. Here is the real shape.
+
+### Every run — automatic, no extra work
+
+- Preflight shows pact, boundary, wiki debt.
+- Free promise re-checks happen before any spending.
+- REFUSE tasks are not dispatched.
+- Ship names the wiki debt and the export staleness.
+
+**Extra time cost: about 10 seconds. Extra token cost: zero** — all of this is
+CLI arithmetic on files ORC already writes.
+
+### Weekly — 2 minutes
+
+```
+$ orc pact status         # any promise drifted?
+$ orc wiki debt           # how far behind is the wiki?
+$ orc export --check      # is AGENTS.md still telling the truth?
+```
+
+All three are free and instant. If they are all green, you are done.
+
+### After a few runs — pay a little
+
+```
+/orc-wiki refresh --top 2        ~$0.20 · 2 minutes
+```
+
+The debt list makes this a small habit instead of a scary $6 event once a
+quarter. **That is the real budget win** — not a cheaper full refresh, but a
+full refresh you never need.
+
+### Monthly — 5 minutes
+
+```
+/orc-aftermath --since 30d       # did last month's work hold up?
+/orc-retro                       # process + result together now
+```
+
+### When something changes shape — rare
+
+```
+/orc-boundary                    # after adding a staging DB, a CI job, a test runner
+/orc-pact                        # after a grill or brainstorm session
+/orc-handoff                     # after a big refactor moves the safe surfaces
+```
+
+---
+
+## 5. Budi's week — the non-developer side
+
+Budi never opens a terminal. He opens the browser panel.
+
+**Monday.** He opens `orc ui` → **Promises**. He reads six sentences about what
+the system guarantees. One says *"refund money goes back to the original payment
+method — exception: store credit, opt-in"*. He now knows, without asking Rina,
+that store credit shipped and how it behaves.
+
+**Tuesday.** Support says the empty-cart message sounds cold. Budi opens
+**Self-serve**, searches "empty cart", edits the English and the Indonesian,
+clicks Preview, then Apply. The check runs and passes. He copies the two git
+commands and sends them to Rina in chat.
+
+Time: 3 minutes. Before: a ticket, and three days.
+
+**Wednesday.** He needs a number for a planning meeting. He opens **Stats →
+Cost** and picks the plan file. The panel has three buttons — Tokens, Quota,
+USD. His team is on Max 20x, so Quota is already selected: store credit used
+**22% of a 5-hour window** and took 27 minutes. He clicks USD to get $7.02 for
+the finance slide. He writes both in the deck.
+
+**Thursday.** He wants a small copy change in an order email. **Self-serve**
+shows it as 🔴 RED with a plain reason: *"looks like text, but it is code — a
+rule inside decides which lines show."* It tells him to ask Rina, and offers to
+save his exact wording as a note so Rina only pastes it. He does that.
+
+He was told no, and it did not feel like a wall.
+
+**Friday.** The **Boundary** panel has one red card: `db/migrations`. He reads
+the three-item checklist. He now understands why the AI does not touch the
+database, in concrete terms, and he can put "add a staging database" on the
+roadmap with a real reason behind it.
+
+**None of this needed Rina.** That is the change.
+
+---
+
+## 6. The compounding — one problem, four lanes
+
+This is the effect that is invisible when you read the lanes one at a time.
+
+In these mocks, one small thing — *"PACT-006 has no test"* — was found four
+different ways:
+
+| Lane | How it found it | When |
+|---|---|---|
+| `/orc-pact` | Refused to mark it HOLDING. Printed `UNCHECKABLE`. | 10-08 |
+| `/orc-boundary` | Marked that task ESCALATE — the agent could not check itself. | 10-08 |
+| `/orc-budget` | Band `[70,80)` ran **151% over**. The executor had nothing to check against, so it guessed and got requeued twice. | 10-08 |
+| `/orc-aftermath` | Still uncheckable 47 days later, and named the other two lanes. | 08-09 |
+
+Four methods. One cause. One cheap fix: write the test.
+
+By the run in this file, the test exists, and:
+- the promise is HOLDING
+- the area is EXECUTE instead of ESCALATE
+- that band's cost went back to forecast
+- and a new store-credit exception could be added **safely**, because there was
+  finally something to prove it against
+
+**One test. Four improvements.** No single lane could have made that case
+convincingly. That agreement is the product.
+
+---
+
+## 7. The money, honestly
+
+Rough monthly numbers for a team doing about 20 runs a month.
+
+### What the new lanes cost
+
+Most of them cost **nothing**, because they are CLI arithmetic on files ORC
+already writes. Only the four that dispatch an agent cost anything.
+
+| Lane | When it runs | Weighted tokens | USD | Quota |
+|---|---|---|---|---|
+| `orc pact status` / `check` | every run + weekly | **0** | **$0** | **0%** |
+| `/orc-pact` reconcile | ~2×/month | 34k each | ~$0.30 | ~1% |
+| `orc boundary status` | every run | **0** | **$0** | **0%** |
+| `/orc-boundary` cards | ~1×/month | 62k | ~$0.60 | ~2% |
+| `orc handoff surfaces` | when asked | **0** | **$0** | **0%** |
+| `/orc-handoff` map | ~1×/month | 27k | ~$0.25 | ~1% |
+| `orc budget forecast/actual` | every run | **0** | **$0** | **0%** |
+| `/orc-aftermath` | monthly | 44k | ~$0.40 | ~1% |
+| `orc export` | weekly | **0** | **$0** | **0%** |
+| **Total added** | | **~201k/month** | **~$2.15** | **~6% of one window** |
+
+### What they save
+
+| Saving | How | Weighted tokens | USD |
+|---|---|---|---|
+| REFUSE tasks never dispatched | ~1 per 3 runs, often in a top band | −740k | −$8 |
+| Fewer repair loops | executors get promises + cards in the slice | −460k | −$5 |
+| Fewer compactions | context risk flagged before the wave, not after | −180k | −$2 |
+| Wiki: light tier on small deltas | ~40% off each refresh | −370k | −$4 |
+| Wiki: never a full refresh | debt cleared in small pieces | −560k | −$6 |
+| Wiki: 2 dead docs retired | never scanned, and never in a slice again | −185k | −$2 |
+| Right lane chosen, with a price | fewer `/orc-ultra` runs "just in case" | −920k | −$10 |
+| **Total saved** | | **~3.4M/month** | **~$37** |
+
+**Net: about 3.2M weighted tokens saved per month, on a 201k investment.**
+In dollars, about **$35 saved on a $2 spend**. On Max 20x, roughly **a whole
+5-hour window given back per month.**
+
+Two notes on reading this:
+
+- **The two wiki lines are the biggest single block** (−1.1M). That is the
+  argument for shipping W1 in the middle of the sequence rather than last.
+- **"Fewer repair loops" is worth more than it looks.** A repair round is almost
+  pure *output* tokens, the most expensive kind. It moves the bill about twice
+  as much as it moves the token count.
+
+And all of this is still the small part. The bigger saving is the reverted run
+in the `/orc-aftermath` mock — one revert costs a developer half a day, which is worth more than
+a whole month of tokens.
+
+---
+
+## 8. What does NOT change
+
+Just as important. If you install all of this:
+
+- **No lane you use today behaves differently by default.** `boundary_gate`
+  ships as `warn`, not `block`. Rina had to turn it up herself.
+- **No new agent is added** except one wiki scanner for the cheap tier.
+- **Nothing runs automatically.** Every paid action is still a thing you ask
+  for. The panel still never runs a lane.
+- **The effort guard is unchanged.** It still matches the exact name `orc`.
+- **`/orc-quick` is still standalone** and still ignores every config key.
+- **No lane blocks you.** Boundary at `block` skips a dispatch — it never stops
+  the run and never refuses your direct instruction.
+- **Traces, run state, checkpoints, resume** — all unchanged. These lanes read
+  them; they do not change how they are written.
+
+---
+
+## 9. Where this could go wrong
+
+An honest list. Each one needs a decision before building.
+
+**1 · Preflight becomes noise.**
+Six lines is already a lot. At ten lines people stop reading, and then a real
+warning is invisible.
+*Rule to hold:* a lane only gets a preflight line when it has something to say.
+`aftermath` in the mock above prints **only** because the area being touched
+churned. On a clean run it prints nothing.
+
+**2 · The pact ledger becomes a second wiki.**
+If every small decision becomes a promise, the ledger grows to 80 entries and
+nobody reads it.
+*Rule to hold:* only `constraint`-tagged decisions become promises, never
+`intent`. And the ledger is reviewed — retirement is offered, not forbidden.
+
+**3 · Boundary at `block` blocks the wrong thing.**
+A wrong REFUSE is worse than no boundary at all, because it stops real work.
+*Rule to hold:* the checklist makes every refusal falsifiable, and `warn` is the
+default for at least one release. Watch `/orc-retro` for tasks that were
+refused and then done by hand with no trouble — that is a mis-calibration
+signal.
+
+**4 · Self-serve writes something bad.**
+The whole safety argument is "a check exists". A weak check that passes is worse
+than no check.
+*Rule to hold:* AMBER exists precisely for this. If ORC is not sure the check is
+strong, the surface is AMBER, and AMBER always shows the human check as a task
+— never as a pass.
+
+**5 · The budget forecast gets trusted too early, or in the wrong unit.**
+A number on a screen looks true. And a dollar figure shown to a Max user is a
+number they will quietly ignore, which is worse than showing nothing.
+*Rules to hold:* `insufficient history` is printed, not hidden. The range is
+never collapsed into one number. Tokens are always available, and cache-read is
+always separable — a blended count re-hides the thing the four-way split exists
+to expose. No dollar figure without a dated price table, no quota figure without
+a known plan, and `unattributed` is always shown.
+
+**6 · The light wiki tier misses something.**
+A cheap scanner writes a shallow doc, and the wiki quietly gets worse.
+*Rule to hold:* `wiki_scan_tier: always_deep` restores today's behaviour
+exactly, the tier is always printed, and STRUCTURAL is always deep. Watch the
+integrity self-check failure rate per tier.
+
+---
+
+## 10. One sentence
+
+**Before:** ORC is a very good build pipeline that forgets everything the moment
+a run ends.
+
+**After:** ORC is a build pipeline that remembers what you promised, knows what
+it should not try, tells you the price first, checks later whether the work
+held up, and lets the person who cannot read code change the things that are
+safe to change.

package/mock-run/context-combiner.md

@@ -0,0 +1,100 @@
+# Mock run — the context combiner
+
+> Two analyses of two related documents, merged into one spec — and it proves
+> that nothing was lost.
+
+There is no slash command for this. It is offered by `/orc-analyze` when two or
+more related analyses already exist.
+
+---
+
+## 1. What it does
+
+You analysed the refund policy PDF last week. Today you analysed the support
+escalation doc. Both talk about refunds. Building from either one alone means
+building half the thing.
+
+The combiner pools **every** requirement from **every** source and gives each
+one exactly one outcome. Then it counts. If the count is not 100%, it does not
+hand off.
+
+---
+
+## 2. The run
+
+```
+/orc-analyze  …  Phase F:
+
+  Two related analyses found. Merge them into one spec?
+  · refund-policy       11 requirements   2026-08-04
+  · support-escalation   7 requirements   2026-08-12
+
+> yes
+```
+
+```
+Dispatching orc-context-combiner-opus-5-high …
+returned as: claude-opus-5 / high        ✓
+
+Phase B — are these really related?
+  shared files       api/refunds/ (both), web/refunds/ (both)
+  shared subject     refund state and who is told about it
+  Verdict: related. Continuing.
+
+Phase C — pooled 18 requirements, reconciled one at a time
+
+  exact duplicate    R-04 = S-02  "refund state must be visible to support"
+                     → ONE row, both sources cited
+  partial overlap    R-07 "notify the customer" vs S-05 "notify support"
+                     → SPLIT into two rows, never collapsed into "notify"
+  conflict           R-09 says refunds are final; S-06 describes a re-open
+                     → CONFLICT, raised to you (below)
+  ordering           S-01 must happen before R-02 (it creates the state)
+
+Phase D — one question at a time
+
+  ❓ R-09 vs S-06 — can a closed refund be re-opened?
+     R-09 (policy PDF, p.4): "a completed refund is final"
+     S-06 (escalation doc, §2): "support may re-open within 14 days"
+     ➡️  The escalation doc is newer. I recommend S-06 wins, with R-09 kept
+        as a note. Your call.
+
+  > S-06 wins, and add the 14 days as a hard rule
+
+Phase E — conservation gate
+  source requirements   18
+  rows in the merged spec 19   (one split into two)
+  every source id has exactly one outcome:
+     kept 15 · merged 2 · split 1→2 · resolved-by-you 1 · dropped 0
+  coverage_pct: 100        → handoff allowed
+
+Phase F — written
+  combined-report.md
+  combined-requirement-spec.md          (a normal spec, plus the source map)
+
+Evidence freshness: 2 anchors from last week no longer match HEAD.
+  api/refunds/create.ts:31 → now :44     (flagged, not silently fixed)
+```
+
+---
+
+## 3. What to notice
+
+- **Nothing may vanish quietly.** Below 100% coverage there is no handoff.
+  Dropping a requirement is allowed — but only as your explicit decision, and
+  it is recorded as one.
+- **A partial overlap is split, never collapsed.** Collapsing two similar
+  requirements into one shorter sentence is how scope disappears.
+- **Old evidence is re-checked against HEAD**, because an anchor that was true
+  last week is not a fact today.
+- **The merged spec is a normal spec.** The planner and the build pipeline do
+  not know or care that it came from two documents.
+- **Full lane only.** There is no mini version of this, and it never builds
+  anything or spawns anything itself.
+
+---
+
+## 4. Related
+
+- Where it starts: [`/orc-analyze`](../templates/skills/orc-analyze/examples/analyze-mock.md)
+- What happens next: [`/orc-plan`](orc-plan.md)