W89 — A Game-Length Life (#376)

Closes #370.

## What was built

The first evidence that the simulation kind survives being played as a
game rather than as a test. Two engine-owned, unpublished regression
fixtures under `src/engine/src/campaigns/long-horizon.ts` (permitted by
`20-contract.md` §19), sharing one job market, course, item, event pair,
opportunity, achievement and headline:

- **`long-horizon-win`** — reaches `goals_met` around week 153-156 via
`player.career.totalWeeksEmployed` hitting 150 and holding two weeks.
- **`long-horizon-loss`** — reaches `failed` via the eviction ladder
(`goalFailurePrecedence: "failure_wins"`), the terminal path no short
fixture had walked before.

`check-content.ts`'s campaign catalogue now imports and registers
`buildLongHorizonWinCampaign` — needed so the content-validation script
can author/build it the way it does every other fixture campaign;
`long-horizon-loss` isn't registered there because it's exercised only
by its own replay test, mirroring the catalogue's existing
one-entry-per-file relationship to other multi-fixture campaign files
(e.g. `stable-life*`). `replay-corpus.ts` needed no edit — it discovers
fixtures by filename prefix scan, so dropping the new
`long-horizon-*.fixture/outcome.json` pairs into `fixtures/replay/` was
sufficient.

Both fixtures' action logs were generated, not hand-authored — a
scripted weekly policy function drove a live engine week by week (via a
temporary, uncommitted generator script), reading current
`SimulationKindState` each week to decide the next plan, exactly like
`design/30-slices.md`'s own W89 sizing argument requires.

## W89.1-W89.8 status

- **W89.1** ✅ `long-horizon-win` plays 150+ weeks to a terminal
`outcome()`, frozen in the corpus like every other entry.
- **W89.2** ✅ Both win and loss committed; the loss ends via the
eviction ladder, not a goal-failure condition.
- **W89.3** ✅ All 27 dispatched `ActionType`s resolve at least once
across the two runs, asserted against `RESOLVER_TABLE`'s non-stub
entries.
- **W89.4** ✅ (with a logged caveat) 14 of the 15 end-of-week systems
are observed acting: 13 non-relationship/non-week_limit systems across
the win+loss runs, plus `week_limit_reached` via an isolated third
side-test campaign (a short scenario with no goals — neither
long-horizon run can reach it without contradicting W89.2's own terminal
paths). `relationships` is excluded by name and logged as a decision: no
weekly relationship rule is specified anywhere in
`10-simulation-kind.md`, so the system can never emit anything beyond
`system.ran`.
- **W89.5** ✅ Byte-identical on replay under the determinism harness,
and byte-identical between a continuous session and a save/restore split
partway through (`engine.serialize`/`.deserialize` mid-run).
- **W89.6** ✅ Serialized size and unbounded-collection counts (week one
vs. final week) are asserted with stated ceilings, not printed.
- **W89.7** ✅ Every previously-committed replay outcome is unchanged
byte-for-byte (verified by the full `npm test` run — all 80 suites /
1193 tests green, including every existing kind's own corpus-match
loop).
- **W89.8** ✅ `npm run typecheck`, `npm run lint`, `npm test` all pass
from `src/engine/`.

## Decisions logged (design/90-decisions.md, mirrored into
docs/docs/engine/OPEN-QUESTIONS.md)

1. **`relationships` stays the simulation kind's only remaining
end-of-week stub.** No weekly relationship rule (decay, drift, or
otherwise) is specified in `10-simulation-kind.md` — §6.11 declares the
state and §7.7 the NPC, but nothing names what a week does to either.
W89.4's coverage assertion counts 14 systems, not 15, excluding it by
name. Revisit when a weekly rule is actually specified (`/contract`
work) — the same call W56's own *Out of scope* already made once.
2. **`long-horizon-loss`'s `pendingEventResponses` grows unbounded** — 0
to 37 entries over its own 160-week run — because nothing in
`endOfWeek.ts`'s `events` system expires an unanswered
`PendingEventResponse`, and the loss run's weekly policy is deliberately
inactive so the eviction arithmetic stays exact. This is a real instance
of the exact defect shape W89 exists to find, found and deliberately
**not fixed** here per W89's own out-of-scope line. W89.6's ceiling
assertion states this as an observed fact of the fixture, not a claim of
boundedness. Revisit by giving `PendingEventResponse` an expiry or a
default-declined resolution — a content/contract decision, not a
slice-sized fix.

Neither is an oversight — both are the interpretations this unit's own
instructions named in advance, applied and recorded rather than
re-litigated.

## Verified

- `cd src/engine && npm run typecheck` — pass.
- `cd src/engine && npm run lint` — pass.
- `cd src/engine && npm test` — pass (80 test files, 1193 tests),
confirming W89.5 and W89.7.
- `./build/ConvertTo-HumanDocumentation.ps1` — run, regenerated
`docs/docs/engine/OPEN-QUESTIONS.md` from the `design/90-decisions.md`
edit (the only marked-block change this PR makes); diff is exactly the
mirrored decision-log entries.
- `./build/Test-Documentation.ps1` — **run to completion, exit code 0.**
Links, terminology, and generated-doc drift all pass.
- `./docs.ps1 -BuildOnly` — not run (Docker-based production build;
relying on the PR's own required *Verify Documentation Build* check for
that gate).

## Still undecided

Both logged decisions above are open follow-ups, not blocking this PR:
the `relationships` system needs a weekly rule specified before it can
be observed, and `pendingEventResponses` needs an
expiry/default-resolution policy before long simulation runs stop
growing that collection unboundedly.

Author: The-Running-Dev

design/90-decisions.md

@@ -314,6 +314,34 @@ fully correct fix, but it touches the shared `applyWorldEffects` interpreter sea
 all six call sites). **Accepted as-is for this MVP slice** — the event is debug severity
 and `scenario` is the only caller reading `.applied` today. Revisit if a second caller
 starts reading `.applied`, or as part of whatever unit closes #349.
+
+**`relationships` (`endOfWeek.ts`) is the simulation kind's only remaining end-of-week
+stub, and W89 is the first thing able to observe that it stays one.** No weekly
+relationship rule — decay, drift, or otherwise — is specified anywhere in
+[`10-simulation-kind.md`](10-simulation-kind.md): §6.11 declares the state and §7.7 the
+NPC, but nothing names what a week does to either, so the system can never emit anything
+beyond `system.ran` (`resolvers.ts`'s own `socialize` is the only thing that moves a
+`RelationshipState`, and only on the player's own action). W89's coverage
+assertion (`long-horizon.replay.test.ts`) therefore reaches fourteen of the fifteen —
+`relationships` is excluded by name rather than silently passed. Thirteen of those
+fourteen are asserted over the two long runs together; `week_limit`, which neither long run
+can reach without contradicting W89.2's two terminal paths, is the fourteenth and has its
+own isolated test in the same file. **Revisit when** a weekly
+relationship rule is actually specified; writing one is `/contract` work, not a slice's
+(W56's own *Out of scope* already made this call once).
+
+**`long-horizon-loss`'s `pendingEventResponses` grows unbounded across its own 160-week
+run — a real instance of the exact defect shape W89 exists to be the first thing able to
+see, found and deliberately not fixed here (W89's own *Out of scope*).** Nothing in
+`endOfWeek.ts`'s `events` system (or anywhere else) expires a `PendingEventResponse` that
+`respond_to_event` never answers; `long-horizon-loss`'s own weekly policy is deliberately
+inactive (`long-horizon.ts`'s header explains why — the eviction-ladder arithmetic has to
+stay exact), so it never answers one, and the collection grows from 0 to 37 entries over
+the run. `long-horizon.replay.test.ts`'s own W89.6 assertion states this as an observed
+ceiling for this fixture, not a claim of boundedness. **Revisit** by giving a
+`PendingEventResponse` some expiry (an `expiresAtWeek`, mirroring `Opportunity`'s own
+field) or an explicit "declined by default" resolution once its `presentWeek` has passed
+by some stated margin — a real content/contract decision, not a slice-sized fix.
 ---
 
 ## 3. Judgement Calls to Revisit (Settled for the MVP)
@@ -1120,3 +1148,25 @@ ledger is one renumbering away from being unexplained.
 Rejected: **Leave it in §19 alone** — the rule is stated and the reasoning is one link away, which is
 defensible today and gets weaker every time the slice ledger is reorganised.
 Reversibility: cheap — documentation only; no code, contract or behaviour change.
+
+### 2026-08-24 — The work mirror's own directory retripped the design-state skip guard it predates
+Context: the 2026-08-21 entry above guarded `tools/Test-DesignState.Tests.ps1`,
+`tools/Read-DesignState.Tests.ps1` and `tools/Update-DesignProjection.Tests.ps1`'s self-referential
+Describe blocks with `-Skip:` computed from whether `design/state/` exists, on the reasoning that this
+repository's compatibility promise (2026-08-19) leaves it unmigrated. `c26b803`/`c8c2e29` (work-mirror
+sync) then started writing `design/state/work/*.md` — a real, intentional adoption of the WorkRef
+record shape, unrelated to full design-state migration — which made the directory exist and flipped
+the guard, un-skipping 9 tests that immediately failed against a repository with no `state-index.md`,
+no `Contract`/`Unit`/`Invariant`/`Decision` records, and no `Check the design state against the tree`
+CI step. `verify.yml`'s "Run Pester tests" step went red on `main` at `7c86a22` as a result; the prior
+session's `.claude/verify-report.json` entry on this branch's predecessor state (commit `8128a0f`)
+already isolated the failures to this cause without fixing it.
+Chosen: point the three guards at `design/state-index.md` instead of the `design/state/` directory —
+the file the self-tests (S16.5) already treat as the migration's own marker, and one `design/state/`
+can no longer answer now that a subdirectory of it has a second, legitimate occupant.
+Rejected: **Exclude `design/state/work/` from the guard's Test-Path instead** — couples a generic
+design-state guard to one specific mirror subdirectory's name, and breaks again the next time
+something else adopts a `design/state/` subpath without adopting the whole migration. **Adopt
+design-state here to satisfy the tests** — the same unscoped-feature-adoption rejection as
+2026-08-21, unchanged by this branch.
+Reversibility: cheap — three `Test-Path` targets, reverted by pointing them back.

docs/docs/engine/OPEN-QUESTIONS.md

@@ -304,6 +304,34 @@ fully correct fix, but it touches the shared `applyWorldEffects` interpreter sea
 all six call sites). **Accepted as-is for this MVP slice** — the event is debug severity
 and `scenario` is the only caller reading `.applied` today. Revisit if a second caller
 starts reading `.applied`, or as part of whatever unit closes #349.
+
+**`relationships` (`endOfWeek.ts`) is the simulation kind's only remaining end-of-week
+stub, and W89 is the first thing able to observe that it stays one.** No weekly
+relationship rule — decay, drift, or otherwise — is specified anywhere in
+[`10-simulation-kind.md`](10-simulation-kind.md): §6.11 declares the state and §7.7 the
+NPC, but nothing names what a week does to either, so the system can never emit anything
+beyond `system.ran` (`resolvers.ts`'s own `socialize` is the only thing that moves a
+`RelationshipState`, and only on the player's own action). W89's coverage
+assertion (`long-horizon.replay.test.ts`) therefore reaches fourteen of the fifteen —
+`relationships` is excluded by name rather than silently passed. Thirteen of those
+fourteen are asserted over the two long runs together; `week_limit`, which neither long run
+can reach without contradicting W89.2's two terminal paths, is the fourteenth and has its
+own isolated test in the same file. **Revisit when** a weekly
+relationship rule is actually specified; writing one is `/contract` work, not a slice's
+(W56's own *Out of scope* already made this call once).
+
+**`long-horizon-loss`'s `pendingEventResponses` grows unbounded across its own 160-week
+run — a real instance of the exact defect shape W89 exists to be the first thing able to
+see, found and deliberately not fixed here (W89's own *Out of scope*).** Nothing in
+`endOfWeek.ts`'s `events` system (or anywhere else) expires a `PendingEventResponse` that
+`respond_to_event` never answers; `long-horizon-loss`'s own weekly policy is deliberately
+inactive (`long-horizon.ts`'s header explains why — the eviction-ladder arithmetic has to
+stay exact), so it never answers one, and the collection grows from 0 to 37 entries over
+the run. `long-horizon.replay.test.ts`'s own W89.6 assertion states this as an observed
+ceiling for this fixture, not a claim of boundedness. **Revisit** by giving a
+`PendingEventResponse` some expiry (an `expiresAtWeek`, mirroring `Opportunity`'s own
+field) or an explicit "declined by default" resolution once its `presentWeek` has passed
+by some stated margin — a real content/contract decision, not a slice-sized fix.
 ---
 
 ## 3. Judgement Calls to Revisit (Settled for the MVP)