Add verified Truthtrance hand inventory questions

Players can ask total-hand and primary-role card-count comparisons using the
existing count controls and compound facts. Component presentation shares the
same default-role classification, including canonical Richese cards, Ix alternate
uses, and Ecaz slot-only specials. Only the respondent receives the verified
aggregate answer; current facts do not impose future custody promises.

Eight new tests cover all 60 catalog cards, real engine actions, malformed input,
private worlds, all four bots, JSON continuation, and a legal custody transfer.
Expanded production-room SQLite cases cover restored total/category/compound
facts, suspended auction state, and one winning concurrent answer/discard.
Independent review's two copy issues were fixed before broad verification.

Validation: types, lint and 4,200 offline tests passed; production build passed;
40 HTTP tests passed. One stable broad run (check 123036 ms, build 6505 ms,
integration 23467 ms). Browser room LC3E55GY verified combined role counts,
total hand size, invalid-input feedback and refresh at 390px phone width.
All 33 pre-integration rooms preserved unchanged; HTTP checks added 30 rooms.
No complete-mode, full-game or publication gate opened.

Private report: 20260913-hand-inventory-final/report.json
Source fingerprint: c6a69a9e69941cdd1e5734d92afedbf10d31fda36c91d26ca7e15f4fb7c59b18
Efficiency pilot milestone 3: aggregate counter 58,695,053 to 59,144,653 at
completed verification = 449,600 tokens, including context/coordination overhead.
Two bounded workhorse-model helpers, reused UI and SQLite fixtures, one browser
room. Counter is not monetary cost and this differently sized batch does not
establish a savings percentage. Final Git/push overhead follows this sample.

docs/TRUTHTRANCE_HAND_INVENTORY.md

@@ -0,0 +1,62 @@
+# Truthtrance hand inventory
+
+13 September 2026. A bounded current-fact capability; public Advanced and unfinished
+expansion starts remain gated. Freeform interpretation and broader commitments are
+still incomplete.
+
+## Contract and authority
+
+The [GF9 November 2020 FAQ, p8](https://www.gf9games.com/dune/wp-content/uploads/2020/11/Dune-FAQ-Nov-2020.pdf)
+permits game-related Yes/No questions and AND/OR combinations, including a weapon
+and defense example. Official-domain indexed page text was retrieved on this date.
+This implementation supplies precise current-hand predicates, not a restriction on
+the questions allowed by the published card.
+
+`HandInventoryFact` compares physical cards currently held with a nonnegative safe
+integer, using exactly, at least or at most. Categories are all cards, primary
+weapons, primary defenses, Worthless, Cheap Hero/Heroine, and remaining Special
+cards. Decks, discards, separate caches and on-table cards are outside the hand.
+Reserved cards still held in hand count. This fact creates no obligation to keep
+cards, acquire more, or use them in a particular slot.
+
+Primary roles are an explicitly worded query definition. They reuse the existing
+component presentation and canonical Richese definitions, not battle-slot legality:
+Weirding Way is primarily a weapon, Chemistry a defense, and Worthless cards remain
+separate even for BG/CHOAM. Stone Burner and Mirror Weapon are printed weapons;
+Portable Snooper is a printed defense despite their persisted `special` kind.
+See [Richese components](RICHESE_COMPONENTS.md) for that physical inventory audit.
+No unfinished effect is enabled by counting its held card.
+
+`game/card-category.ts` provides the shared classification. The focused inventory
+module parses, compares and renders the predicate; the existing Truthtrance flow
+handles priority, aggregate three-valued logic, truthful responses, physical card
+disposal and resuming the interrupted decision. Only the respondent sees the
+verified answer before publication. Exact Yes comparisons necessarily disclose
+the selected count, but matching card identities and individual compound results
+are not added to other views. All four bots use the same owner-only answer.
+
+## Verification and limits
+
+`tests/truthtrance-hand-inventory.test.ts` exercises physical copies, default and
+alternate roles, input rejection, actual questions/answers, private views, four
+profiles, JSON continuations and the absence of a retention promise. The existing
+production-room SQLite suite `tests/truthtrance-spice-recovery.test.ts` now covers
+total/category and compound inventory facts across module restarts, suspended
+auction payment, unchanged private views and competing answers committing once.
+
+Browser acceptance and final required check results are recorded in the milestone
+checkpoint and its private verification report. These checks establish the stated
+fact boundary; they do not certify every card interaction, arbitrary question,
+AI question strategy, complete Advanced play or expansion readiness.
+
+Browser room `LC3E55GY` was created through the lobby, backed up, then staged through
+the genuine Basic initializer into an isolated phase-five question opportunity.
+Its Medium Emperor held four physical cards: two Shields, Lasgun and Baliset.
+The human asker submitted primary-weapon-at-least-one AND
+primary-defense-at-least-two; the AI publicly answered Yes. A negative count
+disabled submission and showed the correction. At 390px width the human then
+submitted total-hand-exactly-four and refreshed during the answer wait. The room
+restored both Yes records, an empty asker hand and the prior movement decision.
+Phone screenshots showed readable controls and the public preview; the temporary
+viewport was reset. Other saved rooms were not used as fixtures. This is targeted
+question/recovery evidence, not a complete-game browser run.