v0.8.2 — every district opens on a Depot, and the docs are pages now

A second-digit bump for a playtest read back against the save file. Nine questions
were asked of one three-Day game; three were bugs, three were the rules working
and undocumented, three were decisions. Every save on the test server was replayed
against this build BEFORE release, which is how the cost of each rule was known
before it was chosen rather than discovered after.

EVERY DISTRICT OPENS ON A DEPOT. A Whistle Post has one A/D track and is not a
Passenger Facility, so the opening of every game was spent unable to work a
passenger and one arrival away from a collision. Two A/D tracks and passengers
from Stage 1 now; "Players start with Whistle Posts, not Depots" is the harder
game, set when the game is created. The deck follows the choice — starting on
Depots the four Depot upgrade cards are left out, because an upgrade must be to
the next tier and a Depot card at a table of Depots is a dead draw. How much
easier it is showed up as a test failure rather than an argument: the cue-coverage
pool needed widening from 24 seeded games to 60 before it held one collision.

NO SAVE WAS STRANDED BY IT, which took care. This is the one house rule that
changes how a game is DEALT rather than how it plays, so replaying a save under
the wrong opening is a different railroad from intent one — silently, with no
error. `withSavedOpening` fills it on the replay paths ONLY. Putting it in the
resolver instead made a fresh Cutthroat game deal Whistle Posts and read as
Custom, which is how the distinction was found.

THREE BUGS, ALL REPORTED FROM ONE GAME AND ALL CONFIRMED ON ITS SAVE.

An Office held TWO TRAINS ON ONE A/D TRACK. The capacity test passed with nothing
standing, the train the Interlocking had been holding at the Limits was moved into
the free slot, and the arriving train was pushed in after it without anyone asking
again whether there was room — so the collision §8.3 calls for never happened. The
held train keeps priority; the newcomer now takes the consequence it would have
met had the held train arrived first.

THE HISTORY FROZE, permanently, and the log cap was not really the cause. Each
seat's "what have I sent you" bookmark was an INDEX into an array the game trims,
so once a seat's bookmark reached the limit the slice returned nothing for the
rest of the game — at a different moment per seat, because each holds its own.
That game's log ended at exactly the cap. Lines carry a sequence number now, which
survives trimming; proven by pushing twice the cap through a simulated seat.

§8.1 ASKED THE WRONG QUESTION TWICE. "Trains may pass" returned `clear` before the
Subdivision was looked at, so a train entering a Double Track was released however
busy the rest of it was — that, not anything about Control Points, is what let
Train 8 out with no ruling. And a train standing at an Office was invisible to the
scan, so one about to re-enter the very Subdivision being entered counted for
nothing. Capacity is the test, not presence: a Depot with a track free is not in
the way; a Whistle Post with its one track taken is.

THINGS THAT HAPPENED SILENTLY NOW SAY SO — a train held against a facing one, a
train released from the Limits (a side effect of somebody else's arrival, so it
simply appeared at the Office), and the train an Interlocking is holding, whose
explanatory tooltip has existed since #99 with NO renderer ever reading the flag.

WHERE A MOVE IS REFUSED, AND WHY. `exploreMoves` decides where the rails go and the
pick-up restrictions are enforced afterwards in `check`, so a square the rails
reached and the card forbade was reachable, un-offered, and absent from the block
list with nothing said. Those squares are blocked with the rule that blocks them
now, and the reasons are got by ASKING `check` rather than re-deriving: a second
implementation of the rules is exactly the failure the block list exists to avoid.
A train may also always recover its own caboose — X13 prints "may drop but not
pick up anything", and a train needs its caboose to be made up, so one that parted
with it could never legally leave again.

RULES DECIDED IN SEPTEMBER AND APPLIED HERE. A Modifier must sit square against its
host, no diagonals. A passenger Modifier may not be played at a Whistle Post. Both
were built, measured, held back for a fortnight so a playtest could finish, and
applied now. A Second Section costs its card: `SECOND_SECTION` was declared in
content.ts and never dealt, so the action was free and the bot ordered 26
accidental ones in a measured round. The card is dealt and spent — gating on a card
the deck never holds would have deleted the mechanic rather than fixed it.

THE DOCUMENTATION IS A SET OF PAGES, not five text files served as text/plain — a
card reference is mostly tables, and as plain text a table is rows of pipes.
Markdown is still the one copy; the build renders it, and publishes the .md beside
each page. No Markdown library: this project has no runtime dependencies and one
would be a poor first. The pages add what Markdown cannot carry without drifting —
a nav across the set, a contents list built from the headings actually rendered,
an anchor on every heading, a 70-character measure, and tables that are tables.
They print as ink on paper.

The references caught up with the rules, checked rather than assumed: two
statements had gone from stale to misleading (the Quickstart told a new player to
"get a Depot down as soon as one appears"), and four rules nobody could look up
are written down — the Office tier table, §8.1 in practice, what the Circus Train
pays for, and that a Realignment can be a card with no legal target.

Adding one card to the deck reshuffles every seeded deal, which broke five
fixtures. Each was a seed meaning "a game like this" — TODO #84, exactly — so
seeds moved and pools widened rather than assertions weakening, and the clearance
fixture pins its terrain the way `enhancements.test.ts` already does. The three
published replays were re-recorded.

Closes TODO #40, #42a, #108, #109 and #110.

1046 fast tests and 35 sim tests pass.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUizFYCMHRWhbWwXhp7WPR

test/replay.test.ts

@@ -22,6 +22,7 @@ import { developerBot, playGame } from '../src/sim/bot.ts';
 import { impediments, isVisible, narrate, phaseLabel } from '../src/sim/narrate.ts';
 import { compress, rehydrateCells, record, renderHtml } from '../src/sim/replay.ts';
 import { summarize } from '../src/sim/stats.ts';
+import { officeProfile } from '../src/engine/content.ts';
 
 // Mirrors `record()`'s own default exactly (`replay.ts`) — "does not drift from the engine" below
 // plays the same seed through both paths and compares outcomes, so they must share one floor.
@@ -32,6 +33,8 @@ const config: GameConfig = {
   maxCollisionsPerDay: DEFAULT_MAX_COLLISIONS_PER_DAY,
   maxCollisionsTotal: DEFAULT_MAX_COLLISIONS_TOTAL,
   pvpCardsAllowed: false,
+  // NO house rules, deliberately: `record()` names none either, so both take today's defaults and
+  // "does not drift from the engine" below compares two games that were dealt the same way.
   optionalRules: {
     reducedVisibility: false,
     employeeRotation: false,
@@ -63,6 +66,7 @@ const SAMPLES: GameEvent[] = [
   // line whose whole job is to be read when nothing happened, so an empty or fallback sentence
   // would reproduce the silence it exists to fix.
   { type: 'freightAgentIdled', player: 0 },
+  { type: 'trainReleasedFromLimits', trainNumber: 8, office: 'Whistle Post', owner: 0, freedBy: 14 },
   { type: 'inboundCleared', player: 0, at: { row: 1, col: 0 }, stock: { type: 'hopper', loaded: true } },
   { type: 'facilityUnjammed', player: 0, at: { row: 1, col: 0 }, from: 'menAtWork', stock: { type: 'hopper', loaded: true } },
   { type: 'trainScheduled', player: 0, trainNumber: 4, roll: 7, slot: 6, rngState: 1 },
@@ -127,7 +131,7 @@ describe('narration', () => {
     /**
      * THE KNOWN GAP, PINNED SO IT CANNOT GROW.
      *
-     * `SAMPLES` exercises the TEXT of 31 of the 56 declared events; the other 25 have a narration
+     * `SAMPLES` exercises the TEXT of 32 of the 57 declared events; the other 25 have a narration
      * case (checked above) but no sample, so nothing proves their sentence is any good. Found
      * 2026-09-09 — the old test built both of its sets from `SAMPLES` and compared them to each
      * other, so it could only ever assert that the sample list had 30 distinct entries, and the one
@@ -217,9 +221,12 @@ describe('impediments', () => {
   });
 
   it('warns when every A/D track is occupied', () => {
-    // Gap 2d — the next arrival is an automatic collision.
+    // Gap 2d — the next arrival is an automatic collision. Filled to CAPACITY rather than to one
+    // train, so the fixture says what it means whatever the Office opens as: a Depot has two A/D
+    // tracks and one occupied is not a warning.
     const s = createGame({ id: 'g', seed: 5, config, playerNames: ['p'] });
-    s.officeAreas.get(0)!.adOccupancy = ['t0'];
+    const area = s.officeAreas.get(0)!;
+    area.adOccupancy = Array.from({ length: officeProfile(area.tier).adTracks }, (_, i) => `t${i}`);
     const found = impediments(s, 0);
     assert.ok(found.some((b) => /A\/D/.test(b.why) && b.severity === 'risk'));
   });
@@ -720,8 +727,14 @@ describe('the replay behaves like the game it is replaying', () => {
      * TWENTY-FOUR, not twelve, and the difference is instructive: on this stride the first game that
      * couples anything is index 13, so a twelve-seed pool still contained none. The rate is what
      * matters, not the count — measured 39 in 200, with couplers at indices 13, 19, 22, 24, 28 …
+     *
+     * SIXTY, not twenty-four, since every player opens on a Depot. Two A/D tracks instead of one is
+     * the single biggest reason a game used to end in a wreck, so collisions went from rare to much
+     * rarer: measured on this stride, no crash cue appears in the first 24 seeds or the first 40,
+     * and the pool needs 60 to contain one. Widened rather than dropped, per the note below — §10
+     * is the one event a player most needs to hear.
      */
-    const recs = Array.from({ length: 24 }, (_, i) => record(1000 + i * 7919, 'standard', 4000));
+    const recs = Array.from({ length: 60 }, (_, i) => record(1000 + i * 7919, 'standard', 4000));
     const withCues = recs.flatMap((rec) => rec.frames.filter((f) => (f.cues?.length ?? 0) > 0));
     assert.ok(withCues.length > 20, `only ${withCues.length} frames carry a cue`);
 
@@ -736,10 +749,10 @@ describe('the replay behaves like the game it is replaying', () => {
     assert.ok(kinds.has('schedule'), 'the 1D12 that sets a train\'s departure Stage landed silently');
     assert.ok(kinds.has('arrive'), 'a train pulling into an Office never made a sound');
     assert.ok(kinds.has('depart'), 'a train highballing out of an Office never made a sound');
-    // Collisions are rare — measured 2 in 40 games — so this is the one cue this pool is not
-    // guaranteed to contain on every stride; it happens to (seeds 96028 and 159380) at the current
-    // stride and seed count. If this starts failing after either changes, widen the pool rather than
-    // deleting the assertion — §10 is the one event a player most needs to hear.
+    // Collisions are rarer still now that every Office opens as a Depot — see the note above. This
+    // is the one cue the pool is not guaranteed to contain on every stride. If it starts failing
+    // after the stride, the seed count or the opening changes, widen the pool rather than deleting
+    // the assertion — §10 is the one event a player most needs to hear.
     assert.ok(kinds.has('crash'), 'a collision never made a sound');
 
     // One CLOCK cue per Stage boundary, the bell replacing the whistle at a Day — the same