The cycle regression gate still read payload.moves, which schema v4
removed in favour of recording.instructions, so every game reported
replay_history_mismatch:0/N and games_with_issues tripped the
--require-zero-issues gate on master (runs 587/594/596).
Verified locally: 12-game gate run reports 0 issues.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PR #170 changed the web replay payload (moves list -> embedded session
recording) but missed these Playwright specs, which only run on master
pushes and so failed post-merge. Assertions now match the v4 shape:
schema_version 4, recording.initial_state present, moves at
recording.instructions.
Verified locally against a real server with freshly built wasm bundles:
full suite 18/18 green, including the five play_canvas specs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two unused-for-free upstream card_game/klondike features:
- GameState now exposes the granular KlondikeStats counters
(move_to_foundation_count, move_to_tableau_count,
move_from_foundation_count, flip_up_count) and the win modal shows a
quiet per-move-type recap line built from them (e.g. "21 to
foundation - 14 tableau moves - 9 flips")
- wasm ReplayPlayer gains seek(step): clamped jump to any position,
rewinding via a stored copy of the recorded deal instead of reparsing
the replay JSON; replay.js Prev now uses it
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replays previously persisted only seed + moves and re-dealt the board
from the seed at playback time, so any change to the seed->deal mapping
(RNG bumps, upstream upgrades) silently invalidated every existing
replay. Schema v4 instead embeds a SessionRecording - the upstream
card_game Session serde ({config, initial_state, instructions}) - so
playback rebuilds the exact recorded board; seed/draw_mode/mode remain
caption metadata only.
- core: SessionRecording newtype delegating to Session<Klondike> serde;
GameState::recording() / from_recording(); from_instructions_unchecked
fixture helper; serde_json added to dev-deps (tests only)
- data: Replay v4 (recording replaces moves); v1-v3 files rejected by
the existing version gate
- engine: win-recording and sync upload freeze game.recording();
playback rebuilds from the recording; Playing carries the extracted
move list (+ Box<Replay> for clippy large_enum_variant)
- wasm: replay_export() builds the full v4 upload payload so JS never
hand-assembles it (the old game.js path hardcoded schema_version: 2
and corrupted u64 seeds via Math.round); ReplayPlayer::from_json
enforces schema_version == 4 with a descriptive error
- web: game.js/play.html use replay_export; replay.js surfaces player
construction errors in the caption instead of dying silently
- server: mode validation accepts data-carrying GameMode variants
(Difficulty uploads previously 400'd against the String field)
Both replays on prod are May-era v1 rows with empty move lists - every
shared replay was already unplayable; the viewer now says why.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The wasm bundles in solitaire_server/web/pkg/ are no longer tracked. The
web-wasm-rebuild workflow (which rebuilt them in CI and committed them
back to master) is gone; instead:
- solitaire_server/Dockerfile gains a wasm-builder stage that runs
build_wasm.sh with the same pinned toolchain (wasm-bindgen 0.2.120,
wasm-pack 0.14.0, binaryen 130) and the runtime image copies pkg/
from it — the image build is now the artifacts' single source of truth.
- web-e2e builds the wasm before Playwright runs, and its trigger paths
now include the wasm-feeding crates it actually tests.
- docker-build triggers on solitaire_data/** and build_wasm.sh too, so
every wasm-affecting change redeploys.
- Self-hosters serving /web or /play from a source checkout run
./build_wasm.sh once (script header and ARCHITECTURE.md updated).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phase 1+2 of the theme-store roadmap: a free catalog served by
solitaire_server and an in-app browse/install flow, making custom
themes installable on Android for the first time (the manual
drop-a-zip flow can't reach the app-private themes dir there).
- solitaire_sync: ThemeCatalogEntry/ThemeCatalogResponse wire types
(additive module; SyncPayload and SyncProvider untouched)
- solitaire_server: THEME_STORE_DIR scan at startup (meta-only
theme.ron parse, sha256, 20 MiB cap, best-effort per archive);
public GET /api/themes, /api/themes/{id}/download, /{id}/preview;
compose volume + README_SERVER docs
- solitaire_data: ThemeStoreClient — catalog fetch + download with
mandatory size/sha256 verification before bytes are released
- solitaire_engine: ThemeStorePlugin — 'Browse theme store' button in
Settings → Cosmetic, modal catalog (leaderboard-style rebuild),
download on AsyncComputeTaskPool, atomic .tmp+rename write into
user_theme_dir, then the existing hardened import_theme pipeline and
an in-place registry refresh
New deps: sha2 (workspace, server+data); ron/zip reused in server;
serde_json added to solitaire_sync dev-deps for DTO round-trip tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Closes the three actionable server Low findings from the 2026-07-06
review:
- #139: unknown-username logins now verify against a static bcrypt
dummy hash so both failure paths pay the same cost — response timing
no longer reveals which usernames exist.
- #140: register maps a unique-constraint violation to UsernameTaken
(409) — the SELECT pre-check stays as the friendly fast path, the
constraint is the arbiter for the concurrent case.
- #141: avatar uploads must start with the magic bytes of the declared
image type; the stored extension (which decides how the file is
re-served) is now backed by content, not the Content-Type header.
Unit tests: real/spoofed/truncated signatures for all four formats,
and the dummy hash's validity at BCRYPT_COST.
Closes#139Closes#140Closes#141
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two concurrency fixes from the 2026-07-06 review (findings M1, M2):
sync push (M1): the load→merge→store cycle ran as three separate DB
operations. Two devices pushing concurrently both read the same stored
payload, merged independently, and the second store overwrote the first
merge — the server visibly regressed until the losing device pushed
again. The whole cycle (including the leaderboard update) now runs in
one transaction; SQLite serialises the writers. The leaderboard
helper's stale docstring (claiming a single conditional UPDATE that was
actually two statements) is corrected — the transaction now provides
the atomicity it described.
refresh rotation (M2): SELECT-then-DELETE let two concurrent refreshes
with the same token both pass the liveness check and both mint fresh
token pairs. The SELECT is gone; rotation now gates on the DELETE's
rows_affected — whoever removes the jti row wins, everyone else gets
401. Sequential reuse was already covered by
consumed_refresh_token_is_rejected, which still passes unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The cycle regression gate did a fresh page.goto() for each of 240 games in one
browser context. Around game ~100 the accumulated resources made
page.waitForFunction time out (30s), failing the gate on ~88% of runs — a
long-standing flaky-CI issue, not a product regression (the 18 e2e tests always
pass).
Load the page once and reset each game via a new __FERROUS_DEBUG__.newGame(seed,
drawThree) bridge method (added to game.js — it was already in play.html).
cycle_metrics.js now navigates once, then loops newGame() + runAutoplay with no
per-game reload, so the run stays fast and stable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The blanket `no-cache` from the earlier fix (#99) regressed the web-e2e cycle
regression gate: it reloads /play-classic 240 times, and with no-cache the
browser re-validated and recompiled the wasm on every load, blowing past the
30s bridge-ready timeout (green at #98, red from #99 onward).
Scope no-cache to just the `include_str!` HTML routes (which change on every
deploy and have no validators — the actual staleness source). The `/web` +
`/assets` ServeDir keep their default Last-Modified caching, so repeated page
loads reuse the downloaded/compiled wasm and the cycle gate is fast again. HTML
freshness — the fix Rhys needed — is unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 2048 surface limit was never wgpu's or the GPU's — it's
downlevel_webgl2_defaults().max_texture_dimension_2d (2048), which
WgpuSettingsPriority::WebGL2 forces. Switch to Functionality: on the WebGL2
(Gl) backend Bevy then adopts the adapter's real limits, which are still
WebGL2-constrained for features/buffers (shaders stay GLES-compatible) but
report the GPU's true max texture dimension (e.g. 16384). The device is
requested with exactly what the adapter offers, so creation can't fail, the
surface is no longer capped, and large viewports (4K) render with no letterbox
and no hardcoded cap.
Removes the resize_constraints cap and the play.html max-width/height caps.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The previous attempts (#97/#98) capped a wrapper element's max-width and
relied on the canvas's width:100% resolving against it — but winit observes
and sizes the *canvas element itself* (via ResizeObserver on its content box),
so the wrapper cap never reached the surface and /play still panicked at
2560x1440 on a 4K@150% viewport.
Use the canonical mechanism instead: set the primary Window's
`resize_constraints { max_width: 2048, max_height: 2048 }`. On web, Bevy maps
this to winit `set_max_inner_size` → the canvas's own `max-width`/`max-height`
style, so the wgpu surface can never exceed wgpu's downlevel_webgl2
max_texture_dimension_2d (2048). play.html mirrors the same `max-*` directly on
#bevy-canvas (belt-and-suspenders) and centres the letterbox with margin:auto;
the obsolete wrapper + clamp script are removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Returning players kept getting stale builds: after the /play canvas fix
deployed, the origin served the new play.html (verified — 8/8 cache-busted
requests) but browsers still rendered the old one even after a hard reload.
Cause: the server sets no Cache-Control on the web router. The HTML pages are
include_str!'d into the binary and the wasm-bindgen output (canvas.js,
canvas_bg.wasm, solitaire_wasm.*) uses fixed filenames that change in place on
every deploy, so browsers heuristically cache them indefinitely.
Add `Cache-Control: no-cache` in the security_headers middleware (which wraps
the web router). Browsers now revalidate before using a cached copy; ServeDir
supplies Last-Modified/ETag so unchanged assets still return a cheap 304, while
changed ones (a new deploy) are re-fetched. Stops the stale-build problem for
everyone going forward.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Correction to the previous canvas clamp (#97), which would NOT have fixed the
crash. It capped the wrapper to the device's gl.MAX_TEXTURE_SIZE, but that's
the hardware limit — on a 4K/integrated GPU it reports 8192+, so the wrapper
was never actually capped and the 2560-wide surface still exceeded the limit.
The real ceiling is wgpu's, not the hardware's: on wasm Bevy creates the device
with Limits::downlevel_webgl2_defaults() (forced by WgpuSettingsPriority::WebGL2
in solitaire_web/src/lib.rs; see bevy_render-0.18.1 settings.rs), whose
max_texture_dimension_2d is a fixed 2048 regardless of GPU. So the cap must be
the constant 2048.
(For the record: the reporter's display is 3840x2160 at 150% scale → a 2560x1440
logical viewport, which is the 2560 in the panic — nothing hardcodes 1440p.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rhys hit a fatal wgpu validation panic loading /play on a 1440p display:
Surface::configure ... Requested was (2560, 1440), maximum extent for
either dimension is 2048
The earlier scale_factor_override(1.0) fix only neutralised HiDPI (CSS×DPR);
it didn't help when the *logical* viewport itself exceeds 2048. `fit_canvas_to_
parent` sizes the wgpu surface to the canvas's parent, so a 2560-wide viewport
configures a 2560-wide surface — past WebGL2's 2048 per-dimension limit on
laptop/integrated GPUs, and the panic kills the WASM thread on the first frame.
Wrap the canvas in #bevy-wrap and, before init, clamp that element's max-
width/height to the device's actual gl.MAX_TEXTURE_SIZE (falling back to the
2048 WebGL2 floor). fit_canvas_to_parent then never produces a surface larger
than the GPU allows. Only devices at the 2048 floor letterbox (centered); GPUs
that report 4096/8192 still fill the viewport.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The /play-classic timer ran at double speed. startTimer() always created a new
setInterval and overwrote timerInterval without clearing the old one, so any
extra call leaked a second interval that also incremented elapsedSecs. The
visibilitychange handler calls startTimer() on "visible", and a load-time
visibilitychange (while startGame's timer is already running) stacks a second
interval — after which stopTimer() only clears one, so the leak persists and
the clock counts ~2x forever.
Guard startTimer() to no-op when an interval is already running. This fixes the
two e2e timer tests that surfaced it once the suite actually ran (they were
asserting 0:03 / 0:04 but seeing 0:06 / 0:13), and the user-visible 2x timer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two latent test bugs surfaced once the e2e suite actually ran (it had been
red on the webServer startup timeout, so these never executed): 14 passed,
4 failed.
- game.js `__FERROUS_DEBUG__` was missing `serialize()` — play.html's bridge
has it but the /play-classic bridge (which the resume/move-history tests
use) drifted. Added it (the wasm SolitaireGame already exposes serialize()).
- game_behaviors.spec.js called `page.clock.tick()`, which is the sinon name;
Playwright's Clock API method is `page.clock.runFor()`. Replaced all 6 calls.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The web-wasm-freshness gate (#93) false-failed because the Bevy wasm build
baked in machine-specific absolute source paths — the cargo registry
(/home/<user>/.cargo/registry/...), the rustup std sources (~/.rustup/...),
and this checkout — so CI's rebuild (under /workspace, /root) never matched the
committed bytes even with identical pinned tool versions.
build_wasm.sh now exports CARGO_ENCODED_RUSTFLAGS with three --remap-path-prefix
rules (cargo home -> /cargo, rustup home -> /rustup, repo -> /build) so the
embedded paths are identical on any machine. It re-states the
getrandom_backend cfg because a *_RUSTFLAGS env var replaces (not merges with)
.cargo/config.toml's target rustflags.
Regenerated all four artifacts with the remap applied — verified zero
machine-specific paths remain (only canonical /cargo + /rustup prefixes), which
is the precondition for the gate's byte-for-byte rebuild-and-diff to match
across machines.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The web-e2e job has been failing on every run with:
Error: Timed out waiting 120000ms from config.webServer.
Playwright's `webServer` runs `cargo run -p solitaire_server --quiet` and
waits 120s for /health. The workflow cached only npm — no Rust cache — so each
run cold-compiled the entire server graph (axum/sqlx/reqwest/aws-lc-sys) inside
that 120s window and never came up. The browser tests never executed; the
failure was infra, not a web regression.
Fix the harness so the tests actually run:
- add `Swatinem/rust-cache@v2` to warm the cargo cache across runs
- add a `Prebuild server` step (`SQLX_OFFLINE=true cargo build -p
solitaire_server`) before Playwright, so `webServer`'s `cargo run` reuses the
compiled binary and starts in seconds (.sqlx offline cache is committed)
- raise `webServer.timeout` 120s -> 300s as a safety margin for a cold cargo
cache (e.g. first run after a deps bump)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The committed `solitaire_server/web/pkg/` artifacts (canvas.* and
solitaire_wasm.*) were last built 2026-06-02 (8b262af) and predate the entire
card_game/klondike migration (#82–#90) plus the v4→v5 save-schema change. The
deployed WASM therefore no longer matched the current source, the JS API glue,
or the HTML (play.html changed in 2cf7282) — which is why the web build was
broken even though the wasm crates compile cleanly on `wasm32-unknown-unknown`.
Rebuilt all four artifacts via build_wasm.sh (wasm-pack 0.14.0,
wasm-bindgen CLI 0.2.120 matching the crate). The regenerated solitaire_wasm.js
and canvas.js glue confirm the API drift (the replay move currency became
KlondikeInstruction in #89).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add [workspace.lints.rust] and wire each member crate up with
[lints] workspace = true:
unsafe_code = "deny" (forbid would break the Android JNI build)
single_use_lifetimes = "warn"
trivial_casts = "warn"
unused_lifetimes = "warn"
unused_qualifications = "warn"
variant_size_differences = "warn"
unexpected_cfgs = "warn"
unsafe_code is "deny" rather than the issue's "forbid" so the three
Android JNI FFI modules (android_keystore, android_clipboard, safe_area)
can opt back in with a scoped #![allow(unsafe_code)] — forbid cannot be
locally overridden. Pure crates carry no unsafe and stay clean.
Clean up the warnings the new lints surface:
- 150ish unused_qualifications removed via `cargo fix` (purely syntactic
redundant-path-prefix removals).
- table_plugin: the TABLE_COLOUR import was #[cfg(test)]-gated while the
camera clear-colour used the fully-qualified path; unqualifying it left
a non-test build with no import. Made the import unconditional instead.
- assets/sources: the `as &[u8]` casts in embed_*_svg! coerce each
fixed-size &[u8; N] to a uniform slice so the tuples fit the
&[(&str, &[u8])] arrays — load-bearing, so scoped #[allow(trivial_casts)].
Workspace clippy -D warnings and the full test suite pass. Android build
not compiled here (needs the NDK; built separately per CLAUDE.md §15) —
the deny + scoped-allow keeps the JNI unsafe blocks legal.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Three independent hardening changes:
1. bcrypt on a blocking thread: hash() and verify() are CPU-bound
(~300 ms at cost 12). Running them directly on an async task starved
the Tokio runtime under concurrent load. Wrapped in spawn_blocking.
2. Async avatar file I/O: std::fs::write/rename/remove_file in an async
handler blocks the executor. Replaced with tokio::fs equivalents.
3. JWT_SECRET minimum length: a secret shorter than 32 bytes is fatally
weak. validate_jwt_secret() now rejects it at startup with a clear
message rather than silently accepting it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
solitaire_wasm/src/lib.rs — 5 new unit tests (9 total, was 4):
- serialize_from_saved_round_trip: board key matches after JSON round-trip
- undo_reverts_to_prior_state: state + history length restored after undo
- draw_one_advances_waste_by_one: DrawOne takes exactly 1 card from stock
- draw_three_advances_waste_by_three: DrawThree takes up to 3 cards
- debug_apply_move_json_stock_click: JSON DebugMove path via native method
solitaire_server/e2e/tests/game_behaviors.spec.js — 5 new Playwright tests:
- resume overlay shows when localStorage save exists; seed() returns null
until user interacts (before bootstrap completes a game)
- clicking New Game on overlay clears history and starts fresh (0 moves)
- clicking Resume restores saved move history length exactly
- HUD new-game button resets history to 0 and score to 0
- tab-visibility timer: timer freezes during hidden, resumes when visible
(tests the visibilitychange fix from the 500-game UX audit); uses
page.clock.install() to control setInterval without real-time delay
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
funman300
funman300