@oneie/claude 0.7.0 → 0.9.0

package/commands/deploy.md

@@ -4,6 +4,8 @@
 
 Ship all five services to Cloudflare. Deterministic sandwich — W0 baseline, build, smoke, approval, parallel deploy, health.
 
+> **Production door (2026-09-05): `bash .claude/scripts/release.sh promote <sha>` then `ship`** — prod ships from `.release/`, a clean checkout carrying a full-suite receipt for its exact tree. `./deploy` stays the pipeline; `release.sh` is what points it at a tree that cannot be dirty. Loop: `../../CLAUDE.md § The dev → prod loop` · plan: `text/release-path-plan.md`.
+>
 > **Monorepo (2026-06-04):** all services live in one repo (`Server/one-ie`). Deploy each via `wrangler deploy` from its folder — **never via repo push, no CI**. Deploys are run locally.
 >
 > **Production cutover (2026-05-23):** Astro site runs on **CF Workers with Static Assets** (not Pages). Production target is **`https://one.ie`** via the `one-prod` worker. Deploy command is **`wrangler deploy` (no `--env` flag)** — see the deploy-target trap note below.
@@ -14,135 +16,377 @@ Ship all five services to Cloudflare. Deterministic sandwich — W0 baseline, bu
 > - **Production (live):** `https://one.ie` — `one-prod` worker, deployed from `one.ie/web/` via bare `wrangler deploy` (**not** `--env production` — see trap note above)
 > - **Gateway (live, stable):** `https://api.one.ie` — `one-gateway` worker
 > - **Pay gateway (live, stable):** `https://pay.one.ie` — `one-core-worker`, deployed from `pay/backend/` via `bun run deploy` (no `--env` flag; no explicit `[[routes]]` block in its wrangler.toml either — same bare-deploy discipline as `one-prod`). Not part of the original "4 services" naming (found 2026-07-05 after a commit there went unshipped for a full deploy cycle) — treat it as a 5th first-class deploy target, not an afterthought.
-> - **Dev (`dev.one.ie` / `one-substrate`) — RETIRED 2026-06-04 (prod-only). NEVER a deploy target.** The sibling tree `apps/dev.one.ie` is dead code — do not build or ship it. The CF git-integration that auto-built `one-substrate` on every push to `one-ie/stage` was disconnected (CF dashboard → Workers & Pages → one-substrate → Settings → Builds). Nothing deploys via repo push anymore.
+> - **Dev (live again since 2026-08-25):** `https://dev.one.ie` — the **`one-dev`** worker, deployed from `one.ie/web/` by `.claude/scripts/deploy-dev.sh` (reachable as **`./deploy dev`**). Its config is *derived from the Astro build output* by `.claude/scripts/gen-dev-config.py` — never hand-rolled — so it cannot drift from what production ships. That script does two things worth knowing: it renames the worker to `one-dev` + binds `dev.one.ie`, and it **strips every cron**. Dev shares production's D1 and KV bindings, so inheriting prod's schedules would mean two workers running the same handlers against the same rows — double sends, double syncs, races. The first dev deploy shipped 6 schedules including `*/5 * * * *` before this was caught. **Dev observes prod data; it must never also drive prod's clock.**
+>   - **`dev.one.ie` is not a data sandbox.** It reads and writes production's rows byte-for-byte. Ship code there freely; treat its DATA as production.
+>   - **The RETIRED thing was `one-substrate`**, the old CF-Pages-built dev project whose git integration was disconnected 2026-06-04. That project stays dead — do not deploy it, and do not confuse it with `one-dev`. The sibling tree `apps/dev.one.ie` is still dead code: do not build or ship it. Nothing deploys via repo push.
 > - **Legacy idle (rollback only — do not deploy):**
 >   - `https://oneie.pages.dev` — old Pages project for `one.ie` ("Ecommerce Playbook" content). Re-attach `one.ie` to this project to roll back.
->   - `app.one.ie`, `demo.one.ie`, `onestudio.dev` — still bound to the prior `one-demo` worker. `one-demo` stays in the account untouched for rollback; redeploying `one-demo` would push stale code, so don't.
+>   - `demo.one.ie`, `onestudio.dev` — still bound to the prior `one-demo` worker. `one-demo` stays in the account untouched for rollback; redeploying `one-demo` would push stale code, so don't.
+>   - `app.one.ie` moved OFF `one-demo` to `one-prod` on 2026-08-06 — it is now an ordinary verified custom domain for `group:one` (D1 `domains` row) and serves `/u/one/*`. Declared in `one.ie/web/wrangler.toml` as a Custom Domain.
 >
 > The custom-domain detach step (Pages → Worker) is done manually via the CF API (no in-repo script today).
 
-## Modes
-
-| Invocation | What |
-|-----------|------|
-| `/deploy` | Full pipeline — W0 + build + 5 services + health |
-| `/deploy astro` | Astro Worker only (re-bundle after UI changes) |
-| `/deploy workers` | Gateway + Sync + Agents only (no Astro rebuild) |
-| `/deploy gateway` | Gateway worker only |
-| `/deploy sync` | Sync worker only |
-| `/deploy agents` | Agents edge agents only |
-| `/deploy pay` | Pay gateway (`one-core-worker` → pay.one.ie) only |
+## Run it — `./deploy`
 
-## The 8-Step Pipeline
+**`.claude/scripts/deploy.sh` is the authority for the procedure. This doc never
+carries a second copy of the steps** — same discipline as `astro.config.mjs §
+vite.ssr.external`. When a step changes, change the script; edit here only when
+what a gate *means* changes.
 
-> **No root orchestrator exists (confirmed 2026-07-08).** There is no root `package.json` and no `scripts/deploy.ts` in this repo — `bun run deploy` / `bun run verify` at repo root **error with "Script not found"**. The commands below are run manually, per-service, from `Server/one-ie`. If a root orchestrator is ever added, update this note; until then, treat every command in this file as the literal thing to run.
-
-**Step 1 — W0 Baseline** (no unified script — run per service)
 ```bash
-# Typecheck all 5 services (each has its own tsconfig, no shared root)
-(cd one.ie/web && bunx tsc --noEmit)
-(cd api && bunx tsc --noEmit)
-(cd sync && bunx tsc --noEmit)
-(cd channels && bunx tsc --noEmit)
-(cd pay/backend && bunx tsc --noEmit)
-
-# Test suite lives only in one.ie/web
-(cd one.ie/web && bunx vitest run)
+./deploy dev            # one.ie/web → one-dev (dev.one.ie) — FAST gate, no approval
+./deploy                # full pipeline — 5 services (PRODUCTION)
+./deploy astro          # one-prod only
+./deploy workers        # api + sync + channels
+./deploy gateway|sync|agents|pay
+./deploy --dry-run      # print every command, ship nothing
 ```
-Record tests passed/total. Fix before proceeding — never deploy on red, **except** entries on the Known-Flaky list below (currently environment-dependent, not code regressions).
 
-**Step 2 — Changes**
-`git diff` summary. Flag large changesets.
+`./deploy` is a repo-root wrapper that `exec`s `.claude/scripts/deploy.sh`;
+either path works, and `/deploy` in Claude Code runs the same script rather than
+retyping its steps.
+
+### The two tiers — one command, two destinations
+
+| | `./deploy dev` | `./deploy` |
+|---|---|---|
+| Destination | `dev.one.ie` (`one-dev`) | `one.ie` + the other four services |
+| Gate | **FAST lane** (`verify:fast`) | **FULL** (`FULL_VERIFY=1`, every test) |
+| Approval prompt | none | yes, unless `--yes` |
+| Migrations | none | `d1 migrations apply --remote` |
+| Crons | stripped | shipped |
+| Who ships it | agents, finishing a loop | a human, promoting |
+
+**`./deploy dev` ships whatever tree it is invoked from — which is `main`.** To
+put a *branch* on dev.one.ie, and to open the PR that proposes it, use
+`bash .claude/scripts/land.sh <branch> --pr --deploy` — one command for
+gate → dev → probe → PR. See "Branch → dev → PR → main" below.
+
+`./deploy dev` exits into `deploy-dev.sh` immediately rather than threading a
+flag through the production pipeline — a different destination with a different
+gate is a different procedure, and `deploy-dev.sh` stays its authority. The one
+production flag that carries over is `--skip-tests`, which becomes the dev
+script's `DEV_SKIP_GATE=1`.
+
+**A fast pass is never reported as a full pass.** Dev going green says the fast
+lane passed and `dev.one.ie` answered 200 — nothing more. Promotion to
+production runs every test again, because the full suite is the gate that has
+actually caught the regressions.
+
+| Mode | Ships |
+|---|---|
+| *(none)* | full pipeline — gates + 5 services + health |
+| `astro` | Astro Worker only (re-bundle after UI changes) |
+| `workers` | Gateway + Sync + Agents (no Astro rebuild) |
+| `gateway` · `sync` · `agents` · `pay` | that one service |
+
+Flags: `--check-creds` (self-test the credential ladder, ship nothing) ·
+`--skip-tests` · `--skip-typecheck` · `--skip-build` · `--skip-migrations`
+· `--skip-health` · `--allow-dirty` · `-y/--yes` (or `DEPLOY_YES=1`) ·
+`-n/--dry-run`. Exit 0 = `deploy:success`; a failed probe exits 1 and prints the
+rollback command. Logs: `.deploy-logs/deploy-<stamp>.log` (gitignored) plus one
+per service for the parallel wave.
+
+### What the script does NOT do — and won't
+
+Two things stay human, by design:
+
+- **Commit / PR / merge** (its own section below). Commit messages and PR bodies
+  need judgment. The script's first gate refuses a dirty tree
+  (`--allow-dirty` overrides) so the deploy and the history can't disagree.
+- **Known-flaky triage.** On a red suite it prints the failing files and stops,
+  pointing at the allowlist. Deciding "that one's the DoH network gap" is a read
+  of the evidence, not a rule. The ONE exception is now mechanised: a suite whose
+  every failure is the shared TypeDB cluster refusing to answer is classified and
+  waived by `typedb-flake-check.sh` — see "The TypeDB flake waiver" below. That
+  check is deliberately narrow; everything else is still your read.
+
+## The gates — what each one asserts
+
+The script runs these in order. Named here so a failure message means something;
+the commands themselves live in the script.
+
+| Gate | Asserts |
+|---|---|
+| **0 · Tree** | working tree clean (`git status --porcelain` empty). First, so a dirty tree fails in a second instead of after a full W0 |
+| **1a · Typecheck** | `bunx tsc --noEmit` clean in all 5 services. Builds `packages/sdk` first when its `dist/` is missing — `one.ie/web` and `channels` resolve SDK types from there, and an unbuilt `dist` fakes a wall of `TS2307` |
+| **1b · Tests** | the full suite in `one.ie/web`, as TWO concurrent lanes — see below. Red blocks, except a classified TypeDB outage; the script won't wave anything else through for you |
+| **3 · Build** | `NODE_ENV=production bun run build` in `one.ie/web` (~30–35s). Emits `dist/server/` + static assets via `@astrojs/cloudflare@13`, patches `wrangler.json` with DO bindings, symlinks `.dev.vars` |
+| **0.4 · Credentials** | Runs **first**, before the slow gates — one curl is cheap, discovering a bad credential after a 2m14s build is not. Resolves ONE credential from an ordered ladder (ambient env → `.env.local` → `one.ie/web/.env` → OAuth), probing each rung with `curl /user` — ground truth, unlike `wrangler whoami`, which answers about whichever credential *that one directory* surfaces. Exports the winner to all five service subshells, so a per-dir `.env` can no longer make two consecutive steps use two different keys. Logs the source as `len=… sha=…`, never bytes, then asserts all five services report the same account id. Self-test: `--check-creds` |
+| **1c · Heavy-gate scheduling** | that vitest and the astro build only overlap when the box can pay for it. Priced in **memory, not cores** — see the section below |
+| **5 · Smoke** | `dist/server/` exists; all 5 `wrangler.toml`s present. Warns if `one.ie/web/wrangler.toml` grew an `[env.production]` block — that's the decoy trap coming back |
+| **6 · Approval** | on `main`, prompts for a literal `yes`. Other branches auto-approve |
+| **6.5 · Migrations** | `wrangler d1 migrations apply DB --remote` — **no `--env`**. "✅ No migrations to apply!" is a pass. Failure blocks: never ship worker code ahead of its schema |
+| **7 · Deploy** | Gateway + Sync + Agents + Pay in parallel (~10s each), then Astro (~30s, largest bundle). Every call bare — no `--env`, ever |
+| **8 · Health** | 4 HTTP probes × 3 tries with backoff, cache-busted with `?_t=`. Sync is cron-only — its clean deploy IS its health signal |
+
+### The suite runs as two lanes, not one run (measured 2026-09-03)
+
+The gate had become a coin flip. Four full runs on an **unchanged tree** gave
+four different failure sets — 8, 12, 13, 8 — and every failing file in all four
+was one of the 19 suites marked `real TypeDB`. Not one of the other ~1126 files
+ever failed. Each run's log carried 41-72 `upstream 503` plus 25-38
+`upstream 500`, and those print only *after* the retry budget is spent
+(`substrate.ts:208`).
+
+The cause is capacity, not code. `substrate.ts:56-60` already recorded it in
+2026-07-25: the shared TypeDB Cloud gateway 503s **in bursts lasting seconds**,
+while the retry cover is 500ms + 1000ms. Eight vitest forks against one external
+singleton turn that into a dice roll — and a dice-roll gate costs far more than
+it saves. The 2026-09-03 deploy spent hours on re-runs and root-cause agents for
+failures that were never in the diff.
+
+`.claude/scripts/test-lanes.sh` splits the suite by its two binding constraints
+and runs them **concurrently**:
+
+| lane | files | bound by | concurrency |
+|---|---|---|---|
+| `pool` | ~1126 | CPU | 8 forks |
+| `typedb` | 19 | one shared gateway | serial (`--no-file-parallelism --maxWorkers=1`) |
+
+The `typedb` lane spends its life waiting on a socket, so it costs almost no CPU
+while `pool` saturates the cores: wall clock is **max(), not sum()**.
+
+| | before | after |
+|---|---|---|
+| wall clock | 238-282s | **121.9s** |
+| result | RED, rotating victims | **BOTH GREEN** — pool 1125 passed, typedb 19 passed |
+| tests | 8-13 failing | 10940 passed |
+
+**This is not `--no-file-parallelism` over the whole suite** — that serialises
+1126 innocent files to fix 19. Every test still runs, no assertion is weakened,
+nothing is skipped; only the suites sharing the external singleton are
+serialised, which removes the contention at its source.
+
+**The selector is read from the tree, never hardcoded.** It greps the
+`real TypeDB` marker the suites already carry, so a new such suite joins the
+serial lane automatically. A frozen list is exactly how this rots back into
+flakiness — `--self-test` asserts the selector finds >0 files, is a strict
+subset, and that every path it names exists.
+
+Both lanes still go through `test-cached.sh`, so passes are memoised per lane
+and a RED is still never cached. `test-full.sh` remains the one definition of
+the vitest flags and hands them over as `TEST_FULL_ARGS`.
 
-**Step 2.5 — Commit, PR, Merge**
-Tests are green at this point, so it is safe to commit. The shared tree is pinned to `main` (`hook:branch-pin`) — never `git checkout`/`switch` there. Do the commit + branch + PR in a scratch worktree instead, then merge via `gh` and fast-forward the shared tree.
 ```bash
-# 1. From the shared tree (still on main): stage by explicit path, never -A
-git status --porcelain   # review scope
-git diff --stat
-
-# 2. Cut a worktree for the deploy branch
-slug="deploy-$(git rev-parse --short HEAD)"
-git worktree add -b "deploy/$slug" ".do-worktrees/$slug" main
-
-# 3. In the worktree: bring over the same paths, commit, push
-cd ".do-worktrees/$slug"
-git add <explicit paths from step 2>
-git commit -m "$(cat <<'EOF'
-deploy: ship <summary of staged changes>
-EOF
-)"
-git push -u origin "deploy/$slug"
-
-# 4. Open PR, wait for CodeRabbit + any checks, merge
-gh pr create --base main --head "deploy/$slug" --title "deploy: <summary>" --body "Automated deploy PR — see /deploy pipeline"
-gh pr merge "deploy/$slug" --squash --auto --delete-branch
-
-# 5. Back in the shared tree: fast-forward and clean up the worktree
-cd /Users/toc/Server/one-ie
-git switch main   # only allowed HEAD move in the shared tree
-git pull origin main
-git worktree remove ".do-worktrees/$slug"
+bash .claude/scripts/test-lanes.sh --list        # show the split, run nothing
+bash .claude/scripts/test-lanes.sh --self-test   # prove the selector still selects
 ```
-Generate the commit message and PR title from the diff — conventional commit format (`feat:`, `fix:`, `chore:`, etc.). If there is nothing to commit (`git status --porcelain` is empty), skip this whole step silently — proceed to Step 3 on the tree as-is.
 
-`gh pr merge --auto` queues the merge and returns immediately if required checks (e.g. CodeRabbit review) are still running; poll `gh pr view "deploy/$slug" --json state,mergedAt` before Step 5's `git pull` if the merge hasn't landed yet — don't build off an unmerged branch.
+**Two lanes mean two `Tests N passed` lines in the log.** `TESTS_REPORT` used to
+`tail -1` and reported only the 19-file lane — the first two-lane production
+deploy said `Tests 143 passed` for a run that executed **10940**. It now sums the
+lanes and says `(2 lanes)`. A gate that is fine while the report understates it
+by two orders of magnitude is the same dishonesty as calling a fast pass a full
+one.
 
-**Step 3 — Build**
-```bash
-cd one.ie/web && NODE_ENV=production bun run build
-```
-Target: ~30-35s. Emits `dist/server/` + static assets via `@astrojs/cloudflare@13`, then patches `wrangler.json` with DO bindings and symlinks `.dev.vars`. Watch for bundle size warnings.
+### Making the deploy faster — two disproved ideas (measured 2026-09-03)
 
-**Fixed 2026-07-19:** the build was V8-heap-OOMing (`FATAL ERROR: Reached heap limit … JavaScript heap out of memory`, `Abort trap: 6`, exit 134) at Node's default ~4 GiB old-space ceiling — not a system memory shortage (5+ GiB free at the time). `package.json`'s `build` script now sets `NODE_OPTIONS=--max-old-space-size=8192` inline, so plain `bun run build` picks it up automatically; no manual env var needed anymore.
+Recorded so nobody spends an afternoon re-deriving them. Both were plausible,
+both are wrong, and the second fails in a way that reads like a pass.
+
+**The gates are at their floor at ~256s. Rescheduling does not move them.**
+From the 2026-09-03 production log:
 
-**Step 4 — Credentials**
-`CLOUDFLARE_API_TOKEN` must be unset. Deploy uses `CLOUDFLARE_GLOBAL_API_KEY` only.
-Scoped tokens lack permissions for workers + custom domains.
-```bash
-unset CLOUDFLARE_API_TOKEN
-echo "${CLOUDFLARE_GLOBAL_API_KEY:+set}" "${CLOUDFLARE_EMAIL:+set}"   # both must print "set"
+```
+vitest  256s   (the lanes, run alone: 121.9s)
+build   256s   (astro's own report: 2m 13s = 133s)
+gates wall-clock: 256s
 ```
 
-**Step 5 — Smoke**
-Verify `dist/server/` exists, all 5 wrangler configs present: `api/wrangler.toml`, `sync/wrangler.toml`, `channels/wrangler.toml`, `one.ie/web/wrangler.toml`, `pay/backend/wrangler.toml`.
+Serial would be 133+122 = 255s — *identical*. Each gate paid ~2x its solo cost
+and the overlap bought nothing, because eight vitest forks plus an 8 GiB-heap
+build on a 10-core box is oversubscription, not parallelism.
 
-**Step 6 — Approval**
-`main` branch: prompts "yes". Other branches: auto-approved.
+The obvious next move — give the build room by capping the pool — makes it
+**worse**:
+
+| | gates wall-clock | build |
+|---|---|---|
+| 8 forks (baseline) | **256s** | 133s |
+| 5 forks (`VERIFY_POOL_FORKS=5`) | **272s** | 209s |
+
+Parallel, serial and capped all land at 250-270s. That is the floor for
+build+suite on this hardware as currently shaped.
+
+**The real target is import cost, and the obvious fix is blocked upstream.**
+The pool lane, run solo, spends more time loading modules than running tests:
 
-**Step 6.5 — D1 Migrations**
-Run before any worker deploy so schema is current when new code lands:
-```bash
-cd one.ie/web && bunx wrangler d1 migrations apply DB --remote
 ```
-**No `--env production`** — same deploy-target trap as the worker deploy itself (see trap note at top of file); `wrangler.toml`'s `[[routes]]`/`[triggers]` were reconciled to top level 2026-07-04 so there is no env-scoped target left to hit, and passing `--env production` here has no matching env block to resolve against.
-"✅ No migrations to apply!" is a pass. Any applied migration is logged and counted (e.g. `0195_workspace_sources.sql` applied 2026-07-08, 4.63ms).
-Migration failures block deploy — never ship worker code ahead of its schema.
+import 145.42s | tests 107.87s
+```
+
+`vitest.config.ts:131` sets `pool: 'forks'`, and every fork re-imports the whole
+module graph independently. Threads share a module cache, so that ought to be
+the win — but **vitest 4.1.7's threads pool is broken in this repo**. Every
+worker dies with `The worker thread was torn down or never initialized. This is
+a bug in Vitest.` Do not reach for `--pool=threads` until vitest is upgraded and
+this is re-tested.
+
+**Read the exit code, not the summary.** That broken run printed:
 
-**Step 7 — Deploy (parallel workers + astro + pay)**
-```bash
-(cd api && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-(cd sync && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-(cd channels && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-(cd pay/backend && unset CLOUDFLARE_API_TOKEN && bun run deploy) &
-wait
-cd one.ie/web && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy   # after the 4 above land
 ```
-Gateway + Sync + Agents + Pay deploy in parallel (~10s each). Astro Worker deploys after (~30s — largest bundle).
+ Test Files  no tests
+      Tests  no tests
+   Duration  162ms
+```
+
+with `rc=1`. "no tests" in a 162ms run is a *collect crash*, one careless glance
+from being reported as a clean pass — the same shape as
+`vitest-collect-crash-reads-as-pass`. An empty selection is never a pass.
+
+**What actually speeds a deploy today:** scope it. `./deploy dev` runs the fast
+lane with no approval; `./deploy workers` skips the astro rebuild entirely when
+`one.ie/web` did not change. The astro build is the long pole and is already
+memoised by tree fingerprint — a warm tree pays none of it.
+
+### Why the gates are not all parallel (measured 2026-08-19)
+
+Steps 1+3 used to launch seven heavy processes with a bare `&`: five `tsc`,
+vitest, and the astro build. That is not parallelism, it is a swap storm. The
+astro build carries `--max-old-space-size=8192` and vitest runs a driver plus
+four 1GB forks, so the two together want ~14GB — on a 24GB box where editors
+and sessions are already resident.
+
+Paging is a cliff, not a slope. Same suite, same commit, same machine:
+
+| vitest ran… | wall-clock | outcome |
+|---|---|---|
+| alone | **176s** | 912 files, 7624 tests, green |
+| beside the build + 5×tsc | **1145s** | 0.0% CPU, no log output, killed |
+
+It happened twice in one day before anyone read it as anything but "tests are
+slow". They are not slow — the suite's own summary accounts for only ~165s of
+actual test time.
+
+**Correction, same day: memory is NOT the proven cause of the hang.** The
+serialisation below is still worth having — 14GB of overlap on a 24GB box is
+real — but the 1145s runs were later reproduced with the build serialised and
+the box at 41% free, no swap thrash, and no network connections held. See
+"The vitest gate hangs" below. Do not cite the memory story as the explanation
+for a hung gate; it explains a *slow* gate, not a *parked* one.
+
+So the heavy gates are now **priced before they launch**: `heavy_free_gb`
+reads free memory the way `lib/govern.sh` does and compares it against
+`DEPLOY_HEAVY_NEED_GB` (default 14). Enough → they overlap as before. Not
+enough → they run one after the other, which is roughly **3x faster
+end-to-end** than "parallel", because the parallel version spends its time
+paging. The five typechecks stay parallel and ungoverned; cached by tree
+fingerprint, they cost ~1s each.
+
+Both heavy gates also run under `gate-run.sh`, for its wall-clock **bound**
+(`DEPLOY_GATE_TIMEOUT`, default 900s — a healthy vitest is 176s and a build
+174s) and its process-group **reap**. macOS ships no `timeout(1)`, and killing
+only the shell reparents the vitest forks to launchd. A hung gate must die on a
+clock rather than outlive the deploy.
+
+A probe that cannot read memory returns a large number, so a broken sensor
+never silently serialises the pipeline. Escape hatches: `DEPLOY_HEAVY_PARALLEL=1`
+forces overlap, `DEPLOY_HEAVY_NEED_GB` retunes the threshold.
+
+### The vitest gate hangs — open, characterised, unexplained
+
+> **Numbers below are from 2026-08-19 and are superseded as measurements** (the
+> suite was 912 files then and is 1145 now; it runs as two lanes since
+> 2026-09-03, ~122s wall). The *diagnosis* still stands and the hang is still
+> unexplained, so the section is kept as-is rather than half-rewritten. Note the
+> non-TTY fork stall below bit again on 2026-09-03: a background re-run with
+> stdout to a file parked until it was killed, and the `script -qeF /dev/null`
+> wrapper fixed it. That wrapper is not optional for any logged run.
+
+
+Three times on 2026-08-19 the vitest gate parked indefinitely and had to be
+killed. What is established:
+
+- The suite itself is healthy: run on its own it completes in **176s**, 912
+  files / 7624 tests / 0 failures, on the same commit and machine.
+- The hang is **not** the astro build competing for memory. It reproduced with
+  the heavy gates serialised, `memory_pressure` reporting 41% free and low
+  pageouts.
+- It is **not** network. At the moment of the hang the vitest main process held
+  **no** open sockets, and neither did its worker.
+- The shape is a fork-pool stall: main parked in `LibuvStreamWrap::OnUvRead`
+  waiting on worker IPC, while its single worker sat at **0.43s CPU / 46MB RSS**
+  seven minutes in — a fork that was spawned and never given work. The log
+  always stops ~90s in, after the jsdom noise, before any test result.
+- `globalSetup` was ruled out: `tests/_global-setup.ts` is local only (reads an
+  env file, prints) with no network call to hang on.
+
+Not established: why. The one difference between every hung run and the green
+one is that the green run passed `--testTimeout/--hookTimeout/--teardownTimeout`
+explicitly — but nothing in that run came close to a timeout, so that may be
+coincidence rather than cause.
+
+Until it is understood, the gate runs under `gate-run.sh`'s bound, so a hang
+now dies on a clock instead of outliving the deploy. If it bites you: the suite
+is trustworthy run directly (`cd one.ie/web && bunx vitest run`), and a deploy
+whose only commits since a green run are harness/doc changes can legitimately
+use `--skip-tests` — verify with
+`git diff --name-only <green-sha>..HEAD | grep -v '^\.claude/'` returning empty.
+
+**Health endpoints** (custom domains only — `*.oneie.workers.dev` is blocked on
+this network, curl exit 6/28): `api.one.ie/health` · `one.ie/api/health`
+(assert `"status":"ok"`) · `channels.one.ie/health` · `pay.one.ie/status`
+(assert `"ok"`; there is no `/health` on that entry — it 404s). The agents
+worker is named **`channels`**, reachable at `channels.one.ie`.
+
+**Build OOM, fixed 2026-07-19:** the build was V8-heap-OOMing (`FATAL ERROR:
+Reached heap limit`, `Abort trap: 6`, exit 134) at Node's default ~4 GiB
+old-space ceiling — not a system memory shortage. `package.json`'s `build`
+script now sets `NODE_OPTIONS=--max-old-space-size=8192` inline; no manual env
+var needed.
+
+**Migrations and the trap:** `wrangler.toml`'s `[[routes]]`/`[triggers]` were
+reconciled to top level 2026-07-04, so there is no env-scoped target left to
+hit and `--env production` has no block to resolve against. Passing it anyway
+silently targets the `one-prod-production` decoy — see the trap note at the top
+of this file.
+
+## Branch → dev → PR → main — `land.sh`, not this doc
+
+**`.claude/scripts/land.sh` is the authority for this procedure**, the same way
+`deploy.sh` is the authority for the pipeline. It replaced the hand-rolled
+`git worktree add` + `gh pr create` recipe that used to sit here; that recipe
+was a second copy of a procedure and it rotted — it still told you to
+`git switch main` in the shared tree, which `hook:branch-pin` refuses.
 
-**Step 8 — Health**
 ```bash
-curl https://api.one.ie/health                   # Gateway
-curl https://one.ie/api/health                   # Astro Worker (production)
-curl https://channels.one.ie/health              # Agents (channels) — custom domain
-curl https://pay.one.ie/                         # Pay gateway — 200 on root is the gate
-curl https://pay.one.ie/status                   # …and /status returns status:"ok" (richer; no /health on this entry)
-# Sync (one-sync) is a CRON-ONLY worker (no HTTP route) — no health endpoint.
-# Its successful `wrangler deploy` ("Deployed one-sync triggers") IS the health signal.
+bash .claude/scripts/land.sh feat/x --pr --deploy --probe / --probe /pricing
 ```
-3 retries with backoff. The 4 HTTP services must return 200. Astro `/api/health` returns `{"status":"ok","agent":"one-prod","model":…,"hasOpenRouter":true}` — assert `status:"ok"`. (The old `units: N` field was removed; there is no `units` to check anymore.)
 
-> **Network gotcha:** `*.oneie.workers.dev` URLs are **blocked on this network** (curl exit 6/28). Always health-check via the custom domains: `api.one.ie`, `one.ie`, `channels.one.ie`. The agents worker is named **`channels`** and is reachable at **`channels.one.ie`**, not `agents.oneie.workers.dev`.
+One command, four steps, in this order:
+
+| # | Step | What it proves |
+|---|---|---|
+| 1 | **gate** — `verify:fast` in the branch's own worktree | the branch is green ON ITS OWN. Not that it survives the trunk — `--pr` deliberately does not merge main in, so the reviewer sees what the branch added and GitHub computes the merge |
+| 2 | **dev** — `$wt/.claude/scripts/deploy-dev.sh` | the branch RUNS. It ships the *worktree's* tree, because `deploy-dev.sh` derives its own ROOT from its own path — `$ROOT`'s copy would ship main and call it the branch |
+| 3 | **probe** — `do-prove.sh`, both bases pinned to dev | the routes answer on dev, under the LANDING RULE. Not `curl / → 200`, which a redirect to `/signin` satisfies |
+| 4 | **PR** — `pr-body.sh` → `gh pr create`/`edit` | a reviewer gets the diff, the trunk drift, the mergeability, the gate label and the dev URL. Re-running UPDATES the open PR, never duplicates it |
+
+Then a human merges the PR, and **promotion to `one.ie` is still `./deploy`** —
+the full gate, from `.release/`. Nothing above touches production.
+
+Four things that are load-bearing, each of which read as a pass before it was fixed:
+
+- **dev.one.ie is one slot, and it writes production's rows.** `--pr --deploy`
+  takes exactly one branch; with two, the second overwrites the first while the
+  first is being probed. Ship code there freely; treat its DATA as production.
+- **Both probe bases are pinned to dev.** `do-prove.sh` falls back to
+  `PROVE_PROD_URL` (default `https://one.ie`) when its dev base is silent — an
+  unreachable dev would otherwise prove *production* and report it as the branch
+  passing.
+- **The probe's route COUNT is read, not its exit code.** `PROVE: skipped (no
+  reachable environment)` exits 0. An unrun probe is not a pass.
+- **The gate is paid once, except under `--quick`.** Step 2 passes
+  `DEV_SKIP_GATE=1` because step 1 just ran `verify:fast` in that same tree.
+  Under `--quick` step 1 was tsc only — not the fast lane — so `deploy-dev.sh`
+  runs its own gate before anything reaches dev.
+
+`land.sh` still has its other two doors: bare (merge main in → gate → `--ff-only`
+main → optionally one dev deploy for the batch) and `--pr` alone (gate → PR, no
+dev). The commit itself stays human — a commit message needs judgment, and
+`deploy.sh`'s first gate refuses a dirty tree so the deploy and the history
+cannot disagree.
 
 ---
 
@@ -267,6 +511,7 @@ Do not remove for production builds.
 | 2026-07-08 | 14.7 MiB | **3.22 MiB** | — | **Exceeds the documented 3 MiB (3072 KiB) free-tier ceiling and still deployed successfully.** Either this account is on a paid Workers plan (10 MiB ceiling) rather than free tier, or the ceiling figure elsewhere in this doc is stale — unconfirmed which. Don't treat "under 3 MiB" as a hard gate until this is resolved; treat 3.2 MiB as the new floor to watch, and re-run Bundle Size Diagnosis if growth continues. |
 | 2026-07-19 | 20.2 MiB | **4.43 MiB** | — | +37% over 2026-07-08's 3.22 MiB. Deployed successfully — no diagnosis run yet. Growth window covers several merged features that day (newsletter platform, directory-submission, social-formats, movers-playbook, etc.) landing in one `/deploy` cycle; not isolated to a single change. Re-run Bundle Size Diagnosis if the next snapshot keeps climbing. |
 | 2026-07-29 | 20.9 MiB | **4.60 MiB** | 1.2 MiB | +6% over 2026-07-19, third consecutive climb. Cheap diagnosis WAS run this cycle (the two grep/`ls` commands below, not a full audit): **no single runaway** — Shiki hits 0, top chunks are `_astro_data-layer-content` 2.1 MiB (content collections), `worker-entry` 1.2 MiB (up from 672 KiB at the 2026-05-22 baseline), `index` 1.3 MiB, `icons` 0.8 MiB, `mermaid` 0.8 MiB, `stripe.esm.worker` 0.6 MiB, `react-vendor` 0.5 MiB. Growth is diffuse feature accretion, not a regression; this deploy's own diff was test-only. Two named candidates if a real audit is ever warranted: `mermaid` (0.8 MiB in the SSR bundle — Rule 2 `ssr.external` candidate if it's only reached from `client:only` islands) and the 15 chunks referencing `react-vendor` (Rule 3 suggests some page still SSR-renders React via `client:load`). |
+| 2026-08-02 | 22.0 MiB | **4.96 MiB** | 1.2 MiB | +8% over 2026-07-29, **fourth consecutive climb**. Cheap diagnosis run again: still **no single runaway** — Shiki 0, `react-vendor` referenced by 15 chunks (unchanged), `worker-entry` flat at 1.2 MiB. Top chunks: `_astro_data-layer-content` 2.1 MiB, `index` 1.3 MiB, **`_broadcast_` 1.3 MiB (new to the top list)**, `worker-entry` 1.2 MiB, `index` 0.9 MiB, `icons` 0.8 MiB, `mermaid` 0.8 MiB, `stripe.esm.worker` 0.6 MiB, `react-vendor` 0.5 MiB, **`generateAuthenticationOptions` 0.5 MiB (new)**. This cycle shipped beautiful-blocks (45 visual blocks + 12 background components), which is a plausible share of the delta. Four climbs in a row with the same "diffuse accretion" verdict each time is itself the signal — the cheap diagnosis has now exhausted what it can tell us, and the two standing candidates (`mermaid` via Rule 2, the 15 `react-vendor` chunks via Rule 3) want a real audit rather than a fifth restatement. |
 
 The 2026-05-22 regression was caused by `build: { inlineStylesheets: 'always' }`
 inlining the full Tailwind stylesheet into every route's manifest entry. One
@@ -276,6 +521,9 @@ char change (`always` → `auto`) saved 8.8 MiB.
 
 ## Service Map (post-migration)
 
+`./deploy <mode>` covers every live row; the per-service command is what the
+script runs, recorded here for rollback and one-off work.
+
 | Service | URL | Config | Deploy command |
 |---------|-----|--------|---------------|
 | Astro Worker (prod) | `one.ie` → `one-prod` | `one.ie/web/wrangler.toml` | `cd one.ie/web && wrangler deploy` — **no `--env production`** (deploy-target trap: appends `-production` to the script name → `one-prod-production`, which nothing routes to; see trap note at top of this file) |
@@ -284,63 +532,79 @@ char change (`always` → `auto`) saved 8.8 MiB.
 | Agents | `channels.one.ie` → `channels` (`*.workers.dev` blocked on this network) | `channels/wrangler.toml` | `cd channels && wrangler deploy` |
 | Pay gateway | `pay.one.ie` → `one-core-worker` | `pay/backend/wrangler.toml` | `cd pay/backend && bun run deploy` (= bare `wrangler deploy`, no `--env` flag) |
 | Pages (legacy idle, rollback) | `oneie.pages.dev` | — | **do not deploy** — rollback target for `one.ie` |
-| Worker (legacy idle, rollback) | `one-demo` (still serves `app.one.ie`, `demo.one.ie`, `onestudio.dev`) | — | **do not deploy** — rollback window for the prod cutover |
+| Worker (legacy idle, rollback) | `one-demo` (still serves `demo.one.ie`, `onestudio.dev`) | — | **do not deploy** — rollback window for the prod cutover |
 
 ---
 
 ## Auth (CRITICAL — never change)
 
-Always: `CLOUDFLARE_GLOBAL_API_KEY` + `CLOUDFLARE_EMAIL`.
 Never: `CLOUDFLARE_API_TOKEN` (scoped token lacks workers + custom domain permissions).
 
+**wrangler reads `CLOUDFLARE_API_KEY` + `CLOUDFLARE_EMAIL`** for global-key auth
+— `CLOUDFLARE_GLOBAL_API_KEY` is *our* name for it and wrangler ignores it. The
+script now exports both spellings off one resolved value, so the two can never
+diverge again.
+
+**The credential lives on disk, not in your shell** — `.env.local` at the repo
+root and `one.ie/web/.env` (same 52-char key). You do not need to export
+anything. If you *do* export one, it wins — which is the trap: a **stale**
+export shadows the good key everywhere, because process env beats wrangler's
+per-directory `.env` autoload. That is exactly how 2026-08-19's deploy used two
+different credentials in two consecutive steps (see the Step 6.6 note below).
+The ladder now probes each rung and falls through a rung that does not answer,
+so a stale export costs a log line instead of a red deploy:
+
+```
+    rejected: ambient env (len=37 sha=eaafbda9) — /user did not answer 200
+  ✓ resolved: global-api-key
+    source: /Users/toc/Server/one-ie/.env.local (len=52 sha=435cba23)
+  ✓ 5/5 services agree on account 627e0c7c…
+```
+
+Ground truth is the API, in **both** directions — it is as able to prove a key
+alive as dead. Never conclude either from wrangler alone:
+
+```bash
+curl -s -o /dev/null -w '%{http_code}\n' \
+  -H "X-Auth-Email: $CLOUDFLARE_EMAIL" -H "X-Auth-Key: $CLOUDFLARE_API_KEY" \
+  https://api.cloudflare.com/client/v4/user     # 200 = fine
+```
+
 The deploy script auto-unsets `CLOUDFLARE_API_TOKEN` from the spawned env to prevent
 accidental use of a scoped token that was exported in the shell.
 
-Required env (export locally before running any deploy step — there is no CI, no root script):
+Required env (export locally before running `./deploy` — there is no CI; the
+script asserts the first two at gate 4 and refuses to ship without them):
 - `CLOUDFLARE_GLOBAL_API_KEY` + `CLOUDFLARE_EMAIL` — auth
 - `PUBLIC_GATEWAY_URL: https://api.one.ie` — build-time-inlined by Astro (**required**; without it the Worker bundle falls back to `one-gateway.oneie.workers.dev` and gateway-backed routes break). Lives in `one.ie/web/.env`.
 
 ---
 
-## Steps
-
-### `/deploy` (full pipeline)
-
-1. Per-service typecheck + `one.ie/web` vitest — W0 gate (see Step 1 above; no unified script exists). Fail here means don't deploy, except Known-Flaky entries.
-2. `git status --porcelain` + `git diff --stat` — surface scope to user. **Stage by explicit path, never `git add -A`** — untracked scratch/backup dirs (e.g. `.wip-backups/`) belong to prior sessions, not this commit.
-3. Commit, PR, merge — worktree branch, push, `gh pr create`, `gh pr merge --squash --auto --delete-branch`, then `git switch main && git pull` on the shared tree (see Step 2.5). Skip entirely if working tree is clean.
-4. `cd one.ie/web && NODE_ENV=production bun run build` — Astro production build.
-5. Verify credentials: assert `CLOUDFLARE_GLOBAL_API_KEY` present, unset `CLOUDFLARE_API_TOKEN`.
-6. Smoke check: assert `dist/server/` exists, all 5 wrangler configs present.
-7. Run D1 migrations (Step 6.5): `bunx wrangler d1 migrations apply DB --remote` — no `--env`.
-8. Parallel deploy: Gateway + Sync + Agents + Pay concurrently, then Astro Worker.
-9. Health checks: 4 HTTP endpoints (Gateway, Astro, Channels, Pay via custom domains, cache-busted with `?_t=$(date +%s)`) × 3 retries with backoff; Sync verified by deploy success.
-10. Report:
-   ```
-   Branch:     main
-   Tests:      4082/4084 pass (2 known-flaky: SSRF DoH lookup, sandbox-network-blocked)
-   Typecheck:  5/5 services clean
-   Build:      33.4s
-   Migrations: 1 applied (0195_workspace_sources.sql)
-   Workers:    gateway+sync+channels+pay parallel, then astro
-   Health:     4/4 HTTP 200 (Gateway, Astro, Channels, Pay) + Sync deploy-confirmed
-   Bundle:     gzip 3220 KiB (exceeds documented 3072 KiB free-tier figure, deploy succeeded — see Verified Bundle Numbers note)
-   ```
-
-### `/deploy astro`
-
-Deploys the production worker (`one-prod`) to `one.ie` from the `one.ie/web/` directory.
-
-1. `cd one.ie/web && NODE_ENV=production bun run build`
-2. Check bundle size: `du -sh one.ie/web/dist/server/`
-3. If > 12 MiB: check which chunk grew (`ls -lhS one.ie/web/dist/server/chunks/ | head -15`)
-4. Run D1 migrations: `cd one.ie/web && bunx wrangler d1 migrations apply DB --remote` (**no `--env production`** — see Step 6.5 note)
-5. `cd one.ie/web && bunx wrangler deploy` — **no `--env production`** (deploy-target trap: see note at top of this file — `--env production` appends `-production` to the worker name and ships to a script nothing routes to)
-6. Health: `curl -sL "https://one.ie/api/health?_t=$(date +%s)"` (expect 200, `status:"ok"`) — always cache-bust with `?_t=`, see Gotchas § cache-control on errors
-
-**Resolved 2026-07-04:** `one.ie/web/package.json`'s `"deploy"` script had the same `--env production` trap. Confirmed via the CF API that it was real — neither `one-prod` (live) nor the `one-prod-production` decoy had any cron schedules registered, meaning `billing-alerts-cron.ts`/`billing-allocation-cron.ts`/`billing-autotopup-cron.ts`/`billing-lifecycle-cron.ts`/`billing-verify-cron.ts`/`funnel-aggregate-cron.ts`/`webhook-deliver.ts`/`broadcast-drain-cron.ts` were not running on any schedule. Fixed by reconciling `wrangler.toml`: the `[[routes]]` (custom domain) and `[triggers]` (crons) blocks — the only two things that existed *only* under `[env.production]` — were moved to top level (everything else was already duplicated there); the now-fully-redundant `[env.production.*]` block was deleted entirely, and `package.json`'s script dropped `--env production`. The trap is now structurally impossible — there's no `--env production` target left to hit.
-
-**Pre-flight (one-time, on cutover only):** ensure no other CF entity owns the `one.ie` custom domain. If wrangler errors with a hostname conflict, detach the prior owner first:
+## Mode-specific notes
+
+Everything common lives in the script. These are the bits that are true of one
+mode only.
+
+### `./deploy astro`
+
+**Resolved 2026-07-04:** `one.ie/web/package.json`'s `"deploy"` script had the
+same `--env production` trap. Confirmed via the CF API that it was real —
+neither `one-prod` (live) nor the `one-prod-production` decoy had any cron
+schedules registered, meaning `billing-alerts-cron.ts` /
+`billing-allocation-cron.ts` / `billing-autotopup-cron.ts` /
+`billing-lifecycle-cron.ts` / `billing-verify-cron.ts` /
+`funnel-aggregate-cron.ts` / `webhook-deliver.ts` / `broadcast-drain-cron.ts`
+were not running on any schedule. Fixed by reconciling `wrangler.toml`: the
+`[[routes]]` (custom domain) and `[triggers]` (crons) blocks — the only two
+things that existed *only* under `[env.production]` — were moved to top level
+(everything else was already duplicated there); the now-fully-redundant
+`[env.production.*]` block was deleted entirely, and `package.json`'s script
+dropped `--env production`. The trap is now structurally impossible — there's
+no `--env production` target left to hit.
+
+**Pre-flight (one-time, on cutover only):** ensure no other CF entity owns the
+`one.ie` custom domain. If wrangler errors with a hostname conflict, detach the
+prior owner first:
 
 ```bash
 # If a Pages project owns it (was `oneie` project pre-cutover):
@@ -350,33 +614,25 @@ curl -s -X DELETE \
   -H "X-Auth-Key: $CLOUDFLARE_GLOBAL_API_KEY"
 ```
 
-### `/deploy workers`
-
-Verified 2026-07-08: all three complete independently with no shared state, so
-run them in parallel (this matches Step 7 of the full pipeline).
-
-```bash
-(cd api && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-(cd sync && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-(cd channels && unset CLOUDFLARE_API_TOKEN && bunx wrangler deploy) &
-wait
-```
-
-Report: version hash per worker, health latency per service.
-
-### `/deploy pay`
+### `./deploy workers`
 
-Deploys the pay gateway (`one-core-worker`) to `pay.one.ie` from the `pay/backend/` directory.
+Verified 2026-07-08: gateway, sync and channels complete independently with no
+shared state — which is why the script runs them concurrently.
 
-1. `cd pay/backend && bunx tsc --noEmit` (no dedicated `typecheck` script — run `tsc` directly)
-2. `unset CLOUDFLARE_API_TOKEN && bun run deploy` (= bare `wrangler deploy`, no `--env` flag)
-3. Health — both probed live 2026-08-02, both **200**, neither authenticated:
-   - Gate: `curl -sL -o /dev/null -w "%{http_code}" https://pay.one.ie/` (expect 200)
-   - Richer: `curl -sL https://pay.one.ie/status` → assert `status:"ok"` (also reports version, destinationMode, contract addresses)
+### `./deploy pay`
 
-   `pay/backend/src/index.ts` mounts `pay/backend/src/routes/status.ts` (`/status`) and `discoveryRoutes` (`/`). It does **not** serve `/health` — the `/health` handler in `pay/backend/src/api/routes/status.ts` belongs to the separate, unmounted `src/api/` tree. Do not health-check `/health` here; it 404s.
+`pay/backend/src/index.ts` mounts `src/routes/status.ts` (`/status`) and
+`discoveryRoutes` (`/`). It does **not** serve `/health` — the `/health`
+handler in `pay/backend/src/api/routes/status.ts` belongs to the separate,
+unmounted `src/api/` tree. Both `/` and `/status` probed live 2026-08-02: 200,
+unauthenticated. `/status` also reports version, destinationMode, and contract
+addresses.
 
-**Found 2026-07-05:** `pay/backend` shipped a real commit (`feat(pay): embeddable payment-link page`) that sat unshipped through a full `/deploy` cycle because the skill's service map only named 4 services. `pay.one.ie` is a first-class 5th deploy target, not a manual afterthought — check `git log` scoped to `pay/` for unshipped commits on every `/deploy` run, the same as the other 4 services.
+**Found 2026-07-05:** `pay/backend` shipped a real commit (`feat(pay):
+embeddable payment-link page`) that sat unshipped through a full deploy cycle
+because the service map only named 4 services. `pay.one.ie` is a first-class
+5th target, not an afterthought — check `git log` scoped to `pay/` for
+unshipped commits, same as the other four.
 
 ## Bundle Size Diagnosis
 
@@ -425,11 +681,63 @@ previous deployment — it needs no worktree and no rebuild.
 
 ---
 
+## The TypeDB flake waiver — a red suite the deploy may ship past
+
+`one.ie/web`'s suite talks to a REAL shared TypeDB Cloud cluster (CLAUDE.md:
+"Don't mock TypeDB in integration tests"). When that cluster blips or a query
+outruns its timeout, a handful of task/substrate suites go red without anything
+in the diff being wrong. That used to be an eyeball judgement, which is exactly
+the call that gets rubber-stamped on the fifth deploy attempt at 2am.
+
+`.claude/scripts/typedb-flake-check.sh` makes it a check. When the vitest gate
+goes red, `deploy.sh` runs it against the gate log:
+
+| exit | means |
+|---|---|
+| 0 | every failure carries a substrate-unavailable signature — waivable, deploy continues |
+| 1 | at least one failure is real, or the log could not be classified — deploy stops |
+
+**It keys on the failure SIGNATURE, never on the filename.** A file-based
+allowlist waives every future failure in that file, including the real ones. The
+waived signatures are `upstream_50[234]`, `status=50[234]`,
+`fixture write failed`, `Test timed out in Nms`, `ETIMEDOUT`, `ECONNRESET`,
+`ECONNREFUSED`, `EAI_AGAIN`, `socket hang up`, `fetch failed`, and TypeDB
+connection errors.
+
+Four properties, each with a red half in `--self-test`:
+
+- **`not_found` is never waivable.** Checked first, independently of everything
+  else. It is the `tasks:claim` privilege boundary and four separate real
+  defects have presented as that exact string. A `not_found` wrapped inside a
+  503 still blocks.
+- **One flake never vouches for its neighbour.** The log is split into vitest's
+  per-failure blocks and EVERY block must carry a signature. A genuine assertion
+  break standing beside a 503 blocks the deploy.
+- **Silence is not a pass.** An empty or unparseable log, or one with no
+  `Tests N failed` line, exits 1.
+- **A waiver is not a green suite.** The report says
+  `WAIVED as TypeDB outage (suite NOT green)`, and a waived run **cannot settle
+  deferred-pin debt** — the waiver speaks only to the failures that reported, not
+  to a pin that never got to.
+
+On by default. `--no-typedb-flake-waiver` (or `DEPLOY_ALLOW_TYPEDB_FLAKE=0`)
+restores the hard stop. Prove the checker still bites before trusting it:
+
+```bash
+bash .claude/scripts/typedb-flake-check.sh --self-test   # 6 cases, 4 of them red halves
+bash .claude/scripts/typedb-flake-check.sh <a-gate-log>
+```
+
+**The waiver is not a diagnosis.** `curl -sS -o /dev/null -w '%{http_code}' https://api.one.ie/health`
+returning 200 while the suite reports 503s means the cluster blipped mid-run. A
+200 alongside failures that are NOT in the signature list means the code is
+wrong — and the checker will tell you so.
+
 ## Known-Flaky Test Allowlist
 
-No allowlist is enforced in code (`scripts/deploy.ts` doesn't exist in this repo — see
-pipeline note above). Treat these by name when they appear in `bunx vitest run` output
-in `one.ie/web`; don't block deploy on them, but don't silently ignore new failures either
+`deploy.sh` deliberately enforces no allowlist — a red suite stops it, and it
+prints the failing files with a pointer here. Triage is yours: treat these by
+name when they appear in `bunx vitest run` output in `one.ie/web`; don't block deploy on them, but don't silently ignore new failures either
 — confirm the failure signature matches before waving it through:
 
 - `tests/e2e/c5-webhook-subscribe.test.ts` — `workflow:webhook-subscribe` "writes a KV
@@ -441,6 +749,19 @@ in `one.ie/web`; don't block deploy on them, but don't silently ignore new failu
   waving through: `curl -v --max-time 5 https://1.1.1.1/dns-query 2>&1 | grep -i refused`
   — if that shows "Connection refused", it's this network gap, not a code regression. If it
   connects fine and the test still fails, it's real — investigate.
+- `tests/unit/tasks-humans.test.ts` + `tests/tasks-do-roundtrip.test.ts` — **only** when
+  the failure is `fixture write failed (status=503 error=upstream_503)` or
+  `(status=502|504 …)`. That message means the shared TypeDB Cloud cluster was
+  unavailable through every retry `typedbQueryDetail` already performs — the substrate
+  refused setup, so nothing downstream proved anything. Verify before waving through:
+  `curl -sS -o /dev/null -w '%{http_code}' https://api.one.ie/health` — a 200 there with a
+  503 in the test means the cluster blipped during the run, not that the code is wrong.
+  **Any OTHER failure in these two files is real and blocks deploy** — in particular a bare
+  `not_found` on `tasks:claim`, which is the privilege boundary and must never be waved
+  through. Four separate defects that used to present as that same `not_found` were fixed
+  2026-08-02 (concurrent-run sweep destruction, silent fixture writes, same-attribute
+  insert races, read-after-write lag); if it reappears, something new is wrong. History:
+  the commit message on `140975e44` and `tests/helpers/probe-sweep.ts`.
 - Hardware/stochastic benchmarks (speed, distribution-timing tests) — expected variance,
   not correctness bugs.
 
@@ -451,8 +772,9 @@ diagnose and fix before proceeding.
 
 ## First-Time Setup
 
-Only needed once. There is no `docs/deploy.md` — this block plus the 8-Step
-Pipeline above is the whole walkthrough. Resource names below are the ones
+Only needed once, before `./deploy` can work at all. There is no
+`docs/deploy.md` — this block plus the script is the whole walkthrough.
+Resource names below are the ones
 actually declared in `one.ie/web/wrangler.toml`; creating differently-named
 resources produces bindings the Worker can't resolve.
 
@@ -475,16 +797,17 @@ bunx wrangler secret put TYPEDB_PASSWORD   # paste at the prompt; never inline t
 cd ..
 
 # First deploy — Worker auto-provisions on first `wrangler deploy`
-# (see "The 8-Step Pipeline" above for the actual per-service commands — no root script)
+# Then just: ./deploy
 ```
 
 ---
 
 ## Logs
 
-No `.deploy.log` / `.deploy-build.log` files are written today — there is no wrapper
-script to produce them (see pipeline note at top of file). Output only lives in each
-`wrangler deploy` command's own stdout; capture it yourself if a durable record is needed.
+`./deploy` writes `.deploy-logs/deploy-<stamp>.log` (gitignored) — every gate's
+stdout in order — plus `.deploy-logs/<service>-<stamp>.log` for each service in
+the parallel wave, since their output would otherwise interleave. The report at
+the end names the log path.
 
 Live logs:
 
@@ -498,6 +821,8 @@ cd one.ie/web && bunx wrangler deployments list --name one-prod | head -10
 
 ## Gotchas
 
+- **A service with no local `wrangler` falls through to a shared `bunx wrangler@latest` cache, and that cache can rot.** Hit 2026-08-06: `sync/` was the only one of the five without wrangler in its `devDependencies`, so `bunx wrangler deploy` resolved to `$TMPDIR/bunx-501-wrangler@latest/` — whose install was missing `esbuild`, so it died `MODULE_NOT_FOUND` before reading a single config. The other four were unaffected because they resolve wrangler from their own `node_modules`. Fixed by pinning `wrangler` into `sync/package.json` like its siblings. If this shape reappears elsewhere, the workaround is a version-pinned invocation (`bunx wrangler@4.80.0 deploy`, which lands in a different cache dir); the fix is a local dep.
+- **Piping `./deploy` into `tail`/`head` masks its exit code** — the pipeline reports the pager's status, not the script's, so a failed run reads as exit 0. `die()` really does `exit 1`; read the ✓/✗ lines or the `.deploy-logs/` file, and don't infer success from a piped exit status.
 - TypeDB Cloud port is **1729** (not 80 or 443)
 - TypeDB HTTP API prefix is `/v1/` (signin, query, databases)
 - Always `CLOUDFLARE_GLOBAL_API_KEY` — scoped tokens lack permissions for workers + custom domains
@@ -510,3 +835,57 @@ cd one.ie/web && bunx wrangler deployments list --name one-prod | head -10
 ---
 
 *Deploy is the closed loop. W0 baseline in, health check out. If health fails, mark() is blocked. Determinism: every step reports numbers, every number gets marked.*
+
+### `code: 9103` does not mean the key was revoked
+
+Measured 2026-08-18, and it cost an hour plus a wrong accusation that a security
+incident had rotated the key. `Unknown X-Auth-Key or X-Auth-Email [code: 9103]`
+at Step 6.5 was a **variable-name mismatch**: the script exported
+`CLOUDFLARE_GLOBAL_API_KEY`, wrangler only reads `CLOUDFLARE_API_KEY`.
+
+Before concluding a Cloudflare credential is dead, ask the API directly — it is
+ground truth and wrangler is not:
+
+```bash
+curl -s -o /dev/null -w '%{http_code}\n' \
+  -H "X-Auth-Email: $CLOUDFLARE_EMAIL" -H "X-Auth-Key: $CLOUDFLARE_API_KEY" \
+  https://api.cloudflare.com/client/v4/user     # 200 = the key is fine
+```
+
+Three companion traps, all real:
+
+- **wrangler 4.x auto-loads `.env` from the cwd**, so a stale key in
+  `one.ie/web/.env` beats both your exported vars and an OAuth session. Isolate a
+  credential test by running it from `/tmp`.
+- **`/user/tokens/verify` is Bearer-only** — `400 Missing "Authorization" header`
+  there is not evidence against a global key. Use `/user` or `/accounts`.
+- **Length proves nothing.** A classic global key is 37 hex chars; a `cfk_`-prefixed
+  one is ~52 and equally valid.
+
+**Never edit `.env` while a deploy is running.** Doing so took a run fully red —
+tsc, vitest, and the astro build — for reasons that had nothing to do with the code.
+
+### The 2026-08-19 sequel: `code: 7403` at Step 6.6, and two keys
+
+The same family, one layer deeper, and worth reading before you diagnose any
+Cloudflare auth failure here. Step 6.5 (D1 in `one.ie/web`) **passed** and Step
+6.6 (D1 in `channels`) **failed** in the same run, seconds apart, on the same
+account. That is only possible if they used different credentials — and they
+did:
+
+- **Three** credentials were reachable from one `./deploy`: a stale 37-char key
+  in the ambient env (injected by `~/.claude/settings.json`'s `env` block — so
+  it existed inside Claude Code sessions and *not* in a plain terminal, which is
+  why grepping the shell profiles found nothing), the good 52-char key in
+  `one.ie/web/.env` + `.env.local`, and the `wrangler login` OAuth session.
+- Nothing *chose* between them. Each service got whatever its own cwd surfaced.
+  Of the five dirs, only `one.ie/web` has a `.env` — so it read the good key and
+  passed, while `channels` fell through to OAuth and hit `7403`.
+- **`7403` ≠ `9103`.** `9103` ("Unknown X-Auth-Key") points at a key; `7403`
+  ("account is not authorized to access this service") points at an account
+  scope, i.e. a session. Reading them as the same symptom is what sent the first
+  diagnosis at a perfectly good key.
+
+Fixed by making the credential *resolved* rather than *ambient* — see Step 0.4
+above. Run `./deploy --check-creds` if you ever doubt which key is in play; it
+names the source and proves all five services agree.