ci(campaigns): every campaign in the pinned content repo builds, or the board is red (task #44)

.github/campaign-build-exclusions.toml

@@ -0,0 +1,39 @@
+# Campaigns in the pinned content repo that `tools/build-every-campaign.py`
+# does NOT require to build, each with the reason it may not.
+#
+# WHY THIS FILE IS NOT A SKIP LIST
+#
+# A silently skipped campaign is a vacuous green (CLAUDE.md: a green gate that
+# binds to nothing is VACUOUS, not a pass) and is worse than no gate at all. So
+# an entry here does not switch a campaign off — it INVERTS the assertion and
+# pins it to a named failure:
+#
+#   * the campaign is still built, on every run;
+#   * it must still fail, and fail with EXACTLY the codes in `expect_codes` —
+#     any other error code is a NEW break that would otherwise have hidden
+#     behind the exclusion, and reds the job;
+#   * if it starts BUILDING, that reds the job too. The exclusion has expired
+#     and must be deleted in the PR that fixed the campaign, or the next
+#     regression it suffers would be excused by a stale line in this file.
+#
+# So the honest states are two, both visible in the job's own output: "built"
+# and "known-red for this reason, still exactly this reason".
+#
+# Fields
+#   id            campaign directory name under <content>/campaigns/
+#   task          the tracking item that will retire this entry
+#   reason        why it cannot build today, in one sentence
+#   expect_codes  the complete set of DW error codes the build must emit
+
+[[exclusion]]
+id = "hollow-vigil"
+task = "#34"
+reason = """
+Two dialogue option labels are authored over the 146 px a vanilla dialog button \
+fits (DW0331); vanilla scrolls an over-wide caption back and forth rather than \
+wrapping it, so the fix is a content rewrite of the labels, not an engine change. \
+Out of scope for the PR that introduced this gate — the gate exists to stop \
+ENGINE changes from breaking campaigns, and hollow-vigil was already red before \
+it.\
+"""
+expect_codes = ["DW0331"]

.github/required-status-checks.txt

@@ -1,6 +1,7 @@
 # The exact `name:` of every CI job that is a REQUIRED status check on `main`
 # (owner decision, 2026-08-05: all ten, previously three; eleven since
-# ADR-0017 added the cross-build shelf).
+# ADR-0017 added the cross-build shelf, twelve since task #44 added the
+# campaign-build gate).
 #
 # THIS FILE IS A DEADLOCK TRIPWIRE. Branch protection matches a required context
 # by its NAME STRING. Rename a job in ci.yml and its required context stops
@@ -30,4 +31,5 @@ skin toolchain (pytest)
 i18n translation tool (pytest)
 mecha cross-check (PR only)
 campaign storybooks (engine-version marker)
+campaign builds (every campaign in the content repo)
 tier 2 (datapack load + PackTest)

.github/workflows/ci.yml

@@ -290,7 +290,7 @@ jobs:
       # Deterministic, offline, stdlib-only python3.
       - name: no early-exit consumer on the right of a pipe (SIGPIPE + pipefail)
         run: python3 tools/check-shell-pipe-shortcircuit.py
-      # All ten jobs are required status checks (owner decision 2026-08-05; it was
+      # EVERY job here is a required status check (owner decision 2026-08-05; it was
       # three, so `tier 2` — datapack load plus the whole PackTest suite — never
       # blocked a merge). Branch protection matches a required context by its NAME
       # STRING: rename a job and its context stops reporting forever, blocking
@@ -362,6 +362,52 @@ jobs:
       - name: every campaign storybook carries a TRUE engine-version marker
         run: python3 tools/check-storybook-version.py
 
+  # Every gate above this line builds a FIXTURE. Nothing in CI had ever built a
+  # real campaign, so the product could stop compiling with the whole board
+  # green: PR #260 reached 10/10 while stopping the flagship released campaign
+  # `nobodys-cave-island` from building at all (26 × `DW0364` on standable cells
+  # at the ocean line). Fixtures exercise one verb each; a campaign is where the
+  # verbs meet a real prefab library, a real layout solve and a real translation
+  # sidecar, and that is the surface an engine change breaks.
+  #
+  # Campaigns are DISCOVERED in the pinned content checkout, never listed here,
+  # so a new campaign on content `main` is gated by the next re-pin with nobody
+  # remembering. A campaign that cannot build today is recorded in
+  # `.github/campaign-build-exclusions.toml`, which INVERTS the assertion instead
+  # of removing it — the campaign is still built and must still fail with exactly
+  # the codes recorded there, so neither a new break hiding behind an exclusion
+  # nor an exclusion that has expired can pass. Every run states its binding
+  # count (CLAUDE.md: a green gate that binds to nothing is VACUOUS).
+  #
+  # Its own job, not a step in `rust`: a campaign build is minutes of layout
+  # solving, and burying it inside the fmt/clippy/test job would make one status
+  # answer two unrelated questions. On every push, not PR-only — a change that
+  # breaks a released campaign must not be able to sit on a green board for the
+  # window between push and PR.
+  campaign-builds:
+    # Name must stay exactly "campaign builds (every campaign in the content
+    # repo)" — it is a required status context in branch protection; renaming it
+    # deadlocks every PR (see .github/required-status-checks.txt).
+    name: campaign builds (every campaign in the content repo)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      # LFS on (the default): a campaign build copies the prefab `.nbt` BODIES
+      # into the datapack, so pointers would fail here for a reason that has
+      # nothing to do with the engine change under test.
+      - uses: ./.github/actions/checkout-content
+      - uses: Swatinem/rust-cache@v2
+        with: { key: campaign-builds }
+      # Built once and handed to the driver, rather than a `cargo run` per
+      # campaign per language: the gate's subject is WHICH engine built the
+      # campaign, so the binary is named explicitly and the script never infers
+      # one. Dev profile — the workspace sets `opt-level = 1` for it (PR #313),
+      # which is what makes a real layout solve affordable on every push.
+      - name: build the engine under test
+        run: cargo build -p delvec --bin delvec
+      - name: every campaign in the pinned content repo still builds
+        run: python3 tools/build-every-campaign.py --delvec target/debug/delvec
+
   # ------------------------------------------------------------------ tier 2 --
   # spec-0004 tier 2 (every PR): build the delve, prove it loads with zero errors
   # on the pinned vanilla server, and run the generated PackTest suite headless.

docs/reference/tools.md

@@ -240,6 +240,7 @@ Never shipped inside a delve.
 |---|---|---|
 | `tools/i18n-translate.py` | agent | `python3 tools/i18n-translate.py <campaign-dir> --lang <code> [--config f] [--delvec cmd] [--batch-size n] [--dry-run] [--force] [--no-validate] [--reflect\|--no-reflect]` — external OpenAI-compatible API, generation-time only; `--reflect` runs the three-step translate → critique → revise pass; see [`i18n.md`](i18n.md) |
 | `tools/skin/` (`delve_skin`) | agent | `python -m delve_skin all <cast.json> --skins-dir D --catalog-dir D --preview-dir D [--id ID] [--scale N]`, or the `build` / `preview` / `catalog` stages individually. Needs its own venv (`pip install -r tools/skin/requirements.txt`); see [`../../tools/skin/README.md`](../../tools/skin/README.md) |
+| `tools/build-every-campaign.py` | CI + agent (run it before proposing any engine change that touches emission, layout or validation) | `python3 tools/build-every-campaign.py --delvec <binary> [--content <checkout>]` — builds **every campaign** the pinned content repo carries, in **every language its `world.json` declares**, and reds if one stops building. Closes the gap that let PR #260 reach 10/10 green while stopping the flagship released campaign `nobodys-cave-island` from building at all (26 × `DW0364`): every other gate builds a FIXTURE, and a fixture exercises one verb, where a campaign is the only place the verbs meet a real prefab library, a real layout solve and a real translation sidecar. Campaigns are **discovered** (any dir under `<content>/campaigns/` with a `world.json`), never listed, so the next content re-pin gates a new campaign with nobody remembering. `--delvec` is required and never inferred — the gate's whole subject is *which engine* built the campaign. A campaign that cannot build today goes in `.github/campaign-build-exclusions.toml`, which **inverts** the assertion rather than removing it: still built, must still fail, and must fail with **exactly** the recorded `expect_codes` — an extra code is a new break that was hiding behind the exclusion, and a SUCCESS is an expired exclusion, both red. Currently one entry: `hollow-vigil`, `DW0331` (task #34). States its binding count every run (discovered / built green / known-red, each named); discovering zero campaigns, building zero campaigns, or an exclusion naming a campaign that no longer exists are each a red. Runs in CI as `campaign builds (every campaign in the content repo)`, on every push |
 | `tools/check-dw-codes.py` | CI | `python3 tools/check-dw-codes.py` — asserts the DW catalog in `compiler.md` matches `crates/**/*.rs` both ways, and that every code has a test |
 | `tools/check-effect-roots.py` | CI | `python3 tools/check-effect-roots.py` — no source file may enumerate the campaign's **effect roots** by hand. An effect root is a `Vec<QuestEffect>` emission can lower; there are five, four hang off the quests stage and the fifth off dialogue, and nothing about the DSL's shape makes them findable by inspection — so every walk that needed "every effect" was written by someone enumerating the roots they knew about. Six were found and fixed independently; a sweep found thirteen more; this gate, on the run that introduced it, found three the sweep had missed (`continuity::excluded_npcs`, `emit::first_damage_players`, `emit_v04_packtests`' despawn scan). **None was ever red** — a walk that visits four of five roots looks correct over any campaign that does not use the fifth. `delvewright_dsl::effects` is now the one enumeration and every walk inherits it, which closes the thirteen; this gate is what stops a fourteenth, since the root fields are ordinary public fields and no type can forbid the loop. Flags a window of 40 source lines naming **3+ distinct** roots, outside an allowlist that carries a REASON per entry (the enumeration itself; `validate::reserved_v06_world`, sound by construction; `plan::required_anchors_for_area`, an open finding printed on every run). **Known non-proof, stated in its docstring**: a proximity heuristic over text — roots spread across a hundred lines, or reached through a helper taking the list as an argument, are invisible. States its binding count every run (currently 128 files, 92 markers); **examining zero files or finding zero markers is a red**, because a renamed field would otherwise leave it quietly green forever. Runs as a step of `docs (local link check)` — a step, not a job, since every job name is a required status context |
 | `tools/check-doc-dupes.py` | CI | `python3 tools/check-doc-dupes.py [path …]` — merge-artifact gate over `docs/**/*.md` + `README.md`: no two body rows in one markdown table share a first-cell key, no heading repeats within a file, no git conflict markers. Kills the class that put `shortcuts[]` in the stage-5 table twice (owner finding 2026-08-03). Same-key rows in *different* tables are fine; a genuine same-table collision means restructure the table, not allowlist it |