Add interactive bidding practice with shared ordinary auction quotes

Expand the internal introduction to eight lessons and six Basic exercises. A Fremen practice auction teaches budget limits, raising, passing and re-entry, full-hand exclusion, all-pass termination, winner-only payment and private purchase inspection. Two explicitly fixed opponent scripts make legal teaching moves without a confirmation step or changes to actual AI strategies.

Extract ordinary bid validation and next-seat/payment/unbid selection into shared pure helpers used by both the practice and live engine. The engine retains actual card custody, ally contribution, Karama, Bureaucrat, payment and response handling. Inconsistent order/eligibility fails closed instead of an unbounded loop. No mode/expansion/publication gate changes.

Version 3 lesson saves retain the existing key and migrate all five v1 and seven v2 positions by identity. Prior committed movement, battle and Traitor choices persist. Only learner bid/pass choices are saved and semantically replayed; over-budget, underbid, full-hand and post-terminal choices are rejected, and unrelated fields are stripped.

Independent review found no material defect. Types, lint, 4925 offline tests and production build pass on stable final source. Nine new tests cover shared bidding order/eligibility/funding and practice outcomes, actual engine payment/card-custody comparisons, all-pass/full-hand paths, all v2 lesson migrations and saved replay. Existing v1, movement, combat and Traitor checks remain green. HTTP/auth/schema code did not change; offline checks include existing in-memory persistence tests, and no live HTTP suite was rerun for this checkpoint.

Browser checks confirm preserved older movement practice, heading focus, keyboard budget control, invalid bid blocking, pass/re-entry, refresh during and after the auction, private Shield inspection, full-hand and all-pass results. The completed bid display stays at the actual committed amount. At 390px controls fit at 309px wide with at least 44px height and no document overflow; the viewport was reset. Screenshot capture still times out, so visual screenshot acceptance remains unverified. The page has no external links and the existing Learn tab is retained.

All 1337 original rooms, 2805 seats and every existing recovery/entry/handover/delegation/migration row are unchanged. SQLite integrity and foreign keys pass. No reset, migration, restart, automation, account-usage check or AI strength calibration occurred. Bounded current-state review also confirmed Extortion, Atomics and Kull still cross previously recorded material rulings; no duplicate questions or house rules were introduced. Alliances, faction/Advanced/expansion teaching and the full game objective remain unfinished.

Verified source tree: b5d318722db155645e98ec3319ffea47f4caa049a06ff5d97abc850db59cb717
This commit is contained in:
Quintine committed 2026-09-20 17:19:18 +10:00
1 parent b9718acda8
commit b7ec94015a
15 files changed
+882 -70

No files matched your search

+36 -8
View File
@@ -4,9 +4,9 @@
from the lobby and rules reference. It adds the previously missing guided
introduction alongside the reference's existing standalone practice wheel.
Seven navigable lessons cover public/private table information and the phase
sequence, shipment budgets, movement, sealed battle plans, Traitor calls, spice
collection and joining or recovering a saved table. Five independent Basic practice positions let a player
Eight navigable lessons cover public/private table information and the phase
sequence, bidding, shipment budgets, movement, sealed battle plans, Traitor calls, spice
collection and joining or recovering a saved table. Six independent Basic practice positions let a player
change inputs, commit, inspect results and retry. These positions are deliberately
separate examples rather than a complete game or an authoritative live room.
@@ -23,6 +23,23 @@ production combat quote. The private Traitor has the existing full inspector.
These fixed teaching scenarios are not AI policy or hidden-information tests
of the four live difficulty levels.
The bidding example gives Fremen six spice and a three-card hand, with a full-hand
variant and two deterministic opponent scripts. Players raise, pass, re-enter on
a later turn or end an unbid auction. The winner pays once; losing bids spend
nothing, and only the player's own purchase exposes a card face/inspector.
Ordinary bid validation and next-seat/all-pass decisions use
`quoteNormalAuctionBid` and `quoteNormalAuctionNext`, extracted from the existing
live engine. The engine retains ally funding, Karama, payment, card custody and
reaction handling. No strategic bot policy changes. Scripted lessons omit those
special interactions and are checked against actual `applyAction` outcomes.
Source authority is the [GF9 base rulebook, Bidding](https://www.gf9games.com/dunegame/wp-content/uploads/Dune-Rulebook.pdf)
and [November 2020 FAQ, p.2](https://www.gf9games.com/dune/wp-content/uploads/2020/11/Dune-FAQ-Nov-2020.pdf#page=2).
The publisher-indexed FAQ text was freshly retrieved on 20 September; it confirms
that a prior passer can bid again. Direct PDF retrieval returned 403, so no new
binary/visual source inspection is claimed. Links remain developer provenance;
the player lesson links only to internal rules.
The movement position starts with five ordinary forces in Red Chasm. It compares
range one with range three from a separate Arrakeen occupant, exact sectors,
storm exclusion and a stronghold already occupied by two other factions. Accepted
@@ -34,10 +51,13 @@ actual authoritative `applyAction` transitions; no live movement rule changed.
The existing wheel, keyboard controls and card inspectors remain available.
Lesson changes focus the heading; responsive columns stack on small screens.
Browser storage contains only a versioned, validated set of lesson choices under
`dune-introduction-v1`. Version 2 payloads explicitly migrate each of the five old
version 1 lesson indices to the same lesson after insertion, preserving every
prior choice and committed outcome. New movement and Traitor fields start at
their defaults; hot updates remount the lesson state to apply this migration.
`dune-introduction-v1`. Version 3 payloads explicitly migrate all five version 1
and seven version 2 lesson positions to the same lesson after insertion,
preserving prior choices and committed outcomes. New bidding fields start at
their defaults; version 1 also receives movement/Traitor defaults. Hot updates
remount the lesson state to apply this migration. Only the player's bid/pass
history is saved; deterministic replay validates the complete auction and rejects
underbids, overspending, full-hand actions and actions after the auction ends.
Future lesson insertions must update the schema and mapping again. The save
includes movement commitments and the before-reveal, revealed and resolved
Traitor stages. Unknown versions,
@@ -59,7 +79,15 @@ and new lobby link stayed within the page bounds. Screenshot capture timed out,
so visual screenshot acceptance remains open; DOM bounds are narrower evidence.
Final source-bound reports and Git delivery record required check/build results.
This does not replace a full interactive curriculum. Bidding, alliances, broader movement/arrival powers, faction lessons, Advanced
`tests/introduction-bidding.test.ts` compares offered openings and re-entry with
actual engine payment and physical card custody, checks all-pass/full-hand paths,
all seven version 2 lesson migrations and exact saved bidding continuation.
`tests/normal-auction.test.ts` exercises shared turn order, eligibility, pass
re-entry, finite termination and integer/funding limits. Final source-bound
browser/check/build and preservation evidence for bidding belongs to its private
checkpoint and Git message.
This does not replace a full interactive curriculum. Alliances, broader movement/arrival powers, faction lessons, Advanced
support and expansion lessons remain future work. The expanded lessons retain
Partial/Prototyped coverage. Final browser/check/build, preservation and Git
evidence for this follow-up belong to its source-bound private checkpoint. No rules mode, expansion or publication gate is opened.