Every playthrough showed the exact same fixed 30 letters in the exact
same order — no randomization anywhere in the codebase. Added 15 new
letters (3 per day, l31-l45), bringing each day's pool to 9, and
changed the engine to draw PER_DAY_COUNT (6) at random from that pool
per game instead of always using the same fixed set.
- engine.ts: buildRoster() shuffles each day's pool and takes the
first 6; GameState now carries a `roster` (letter ids in day order)
generated once at createState() and persisted with the rest of the
save. lettersForDay/currentLetter read through the roster instead
of the full LETTERS array directly.
- Save key bumped to v3 — old saves have no roster field, and
backfilling one wouldn't match already-recorded decisions, so old
saves are simply ignored rather than partially migrated.
- Tests rewritten: several previously asserted specific letter ids at
specific positions, which no longer holds under randomization.
Split into buildRoster() property tests (right count, no dupes,
every id resolves, different seeds differ) and engine tests that
either check general correctness or construct an explicit roster
when a specific letter's behavior needs locking in (l22, l04).
Verified: 11/11 vitest, tsc clean, and a real 5-fresh-games browser
test confirming 5 different first-letter senders instead of the same
one every time.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Letters l01-l20 only ever hit the generic default consequence lines
(engine.ts's DEFAULT_WRONG_CONSEQUENCES / a generic correct-route
line) — l21-l30 already had the good stuff. Gave all 20 their own
correctConsequence and wrongConsequences, matching the dry,
institutional dread already established, tied to each letter's own
specific details (the cardigan, the wrong key, the room removed from
the blueprint, Avdelning Noll's negative floors, the future self
warning about 16:42). Bumped a few penalty overrides (l04, l06, l08,
l09, l13, l15, l18) where the wrong choice is narratively worse than
the default 1/1 cost, same pattern already used on l22/l26.
Also split the single win/loss ending into four tiers based on final
suspicion: a clean TILLSVIDARE vs. a "TILLSVIDARE, VILLKORLIGT" win,
and an OMPLACERAD (composure-driven) vs. AVFÖRD (suspicion-maxed)
loss — same stamp-and-personnel-file frame, different outcome text,
so a second playthrough has something new to see even on the same
ending type.
Verified: tsc clean, 8/8 vitest (added a data-integrity check that
every letter has a correctConsequence, plus a test locking in l04's
new penalty), a real browser playthrough of l01's new consequence
text, and all four ending variants rendered by injecting finished
game states directly rather than scripting a fragile 30-letter
keyboard sequence.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
