# ACCESSIBILITY SPECIFICATION **Phase 2 contract.** Written 2026-08-09. Governing prompt: `md/06_HUD_INTERACTION_FEEDBACK_ACCESSIBILITY.md` §6. **The instruction being executed.** *"Explore without changing default Classic mechanics … For each accessibility option document whether it is presentation-only / input-equivalent / mechanical assist / score-invalidating assist. If an assist changes challenge, label it."* **The stance this document takes.** Defender is a hard game on purpose, and its difficulty is part of what is being preserved. Accessibility work here is therefore split cleanly: everything that removes a *barrier to perceiving or operating* the game is offered freely, on by default where the platform signals it, and costs the player nothing. Everything that removes a *barrier to succeeding* is offered too — because a game nobody can finish is not preserved either — but it is labeled, it is recorded in the session metadata, and it does not enter the same score table. Nobody is excluded and nobody is deceived. **Binding upstream.** `ARCHITECTURE.md`, `HUD_RADAR_SPEC.md`, `INTERACTION_FEEDBACK_SPEC.md`, `DISPLAY_GEOMETRY.md`. --- ## 1. The four classes, defined precisely The classification is not a vibe. Each class has a mechanical test, and every option in §3 has been run against it. ### Class P — **Presentation-only** > The simulation is bit-identical. Running the same `{seed, inputSequence, tickCount}` with the > option on and off produces identical state at every tick. **Test:** `assertReplayIdentical(replay, {option: on}, {option: off})` — identical score-event stream, identical per-frame entity counts, identical final RNG state. **Score effect:** none. Scores are fully comparable. Replays are interchangeable. **Label in UI:** none required. ### Class I — **Input-equivalent** > The simulation is untouched, and the option changes only how the seven boolean input lines are > produced from the user's hardware. **The produced sequence of input states must lie inside the set > of sequences a cabinet's switches could have produced.** That last clause is the whole test, and it excludes more than it first appears. A *hold-to-toggle* converter for thrust is Class I, because the resulting thrust line is still an arbitrary boolean the cabinet's toggle-switch-shaped brain could have produced. **Auto-fire is not**, because it produces a fire-line pattern at a rate no human hand on a cabinet button could sustain, and firing rate is a difficulty parameter. **Test:** with the option on, record the seven-line input trace; assert every state transition is attributable to a user action or to a bounded state machine over user actions, with no timer-driven transitions at a rate above 10 Hz. **Score effect:** none. Scores comparable. Replays interchangeable (a replay stores the *input lines*, not the key events, so the converter is upstream of the recording). **Label in UI:** none required. ### Class M — **Mechanical assist** > The simulation behaves differently. A rule, a rate, a constant or a gate has changed. **Score effect:** the run is **flagged**. It is scored, it is shown, it goes into a separate "Assisted" table, and its metadata names every assist that was active. It is not deleted and it is not sneered at. **Label in UI:** required, at three points — in the settings row itself, as a persistent marker in the status band while active, and on the score screen. Wording is specified per option. ### Class S — **Score-invalidating assist** > The simulation behaves differently in a way that removes the run's meaning as a Defender > performance: unlimited resources, invulnerability, time manipulation, or direct state editing. **Score effect:** **no score is recorded at all.** The Hall of Fame is not offered. Replay metadata carries `scored: false` and the reason. **Label in UI:** required, and the status-band marker is persistent and unmissable. ### The one-way ratchet **Enabling a Class M or S option mid-run immediately and permanently downgrades that run.** Turning it back off does not restore the run's status. This is the only way the classification means anything. A run's class is `max(class of every option enabled at any point during the run)`. --- ## 2. Platform defaults — what is on before the player touches anything | Signal | Effect | Class | |---|---|---| | `prefers-reduced-motion: reduce` | Screen shake off (`shakeScale = 0`); page-transition animations off; the shell's own motion disabled. **The game's own motion — the ship, the camera ease, the scanner cadence — is untouched**, because it is the game. | P | | `prefers-contrast: more` | High-contrast scanner palette (A-08) on; shell UI contrast raised. | P | | `prefers-color-scheme` | Shell chrome only. The game canvas is always black-surrounded (`DISPLAY_GEOMETRY.md` §5.1 rule 5). | P | | `prefers-reduced-transparency` | Shell overlays become opaque. | P | | Screen reader detected (heuristic: focus moves without pointer events) | Nothing is auto-enabled. **We do not guess at disability.** The shell is fully navigable and announced regardless; see §6. | — | **Nothing in Class M or S is ever enabled by a platform signal.** A media query cannot consent to a difficulty change. --- ## 3. The option matrix Every option from `md/06` §6, in its order, plus the ones the analysis showed to be missing. | ID | Option | Class | Default | Label required | |---|---|---|---|---| | A-01 | Full remapping (keyboard, gamepad) | **I** | on | no | | A-02 | Sticky / hold-alternative for Thrust | **I** | off | no | | A-02b | Sticky / hold-alternative for Up and Down | **I** | off | no | | A-03 | Auto-fire | **M** | off | **yes** | | A-04 | One-button mode (single key cycles-and-fires) | **S** | off | **yes** | | A-05 | Reduced flash | **P** | off (on if reduced-motion) | no | | A-06 | Reduced bloom | **P** | off | no | | A-07 | Reduced screen shake | **P** | off (on if reduced-motion) | no | | A-08 | High-contrast scanner | **P** | off (on if prefers-contrast) | no | | A-09 | Color-vision palettes (4 variants) | **P** | off | no | | A-10 | Larger HUD (1×, 1.5×, 2×) | **P** | 1× | no | | A-11 | Audio cue reinforcement | **P** | off | no | | A-12 | Captions / event text | **P** | off (one caption on by default in Enhanced) | no | | A-13 | Haptic cue reinforcement | **P** | on when a gamepad is present | no | | A-14 | Pause-friendly practice mode | **S** | off | **yes** | | A-15 | Configurable hyperspace flashing | **P** | standard | no | | A-16 | Photosensitivity-safe presentation | **P** | off | no | | A-17 | Slow motion (0.75× / 0.5× / 0.25×) | **S** | off | **yes** | | A-18 | Extra lives (3 / 5 / 9) | **M** | 3 (factory) | **yes** if ≠ factory | | A-19 | Invulnerability | **S** | off | **yes** | | A-20 | Hyperspace never kills | **M** | off | **yes** | | A-21 | Scanner smoothing (60 Hz blips) | **P** | off | **yes**¹ | | A-22 | Larger blips | **P** | off | no | | A-23 | Input latency compensation | — | **not offered** | — | | A-24 | Aim assist / target lock | — | **not offered** | — | | A-25 | Shell keyboard navigation & screen-reader support | **P** | always on | no | | A-26 | Shell text scaling & reflow | **P** | follows the browser | no | | A-27 | Audio: master / effects / drone levels, mono downmix | **P** | 100 / 100 / 100, stereo | no | | A-28 | Pause on focus loss | **P** | on | no | | A-29 | Confirm-before-hyperspace | **I** | off | no | ¹ A-21 is Class P by the mechanical test — the core is untouched — but it is labeled anyway, because a veteran will feel it and is entitled to be told. See `HUD_RADAR_SPEC.md` §9.3. This is the only Class P option that carries a label, and the exception is deliberate. --- ## 4. Per-option specification --- ### A-01 · Full remapping — **Class I**, default on Fully specified in `INTERACTION_FEEDBACK_SPEC.md` §6.5. Summary of the accessibility-relevant properties: - Bindings key off `KeyboardEvent.code` (physical position), labels come from `navigator.keyboard.getLayoutMap()`. A player on any layout gets the same hand shape and correct labels. - Up to **3 bindings per action**, so a player can bind both a key and a gamepad button and use whichever hand is available that day. - **Any action may be left unbound** — a warning, never an error. - **Modifier-only bindings are permitted** (`LeftShift`, `RightAlt`) for players who cannot reach the letter keys. - No chord bindings. A chord is a two-finger requirement, which is exactly the barrier remapping exists to remove. - Remapping is reachable **from the pause menu without leaving the run** and takes effect on the next frame. **Why Class I:** the six switch lines are produced from different physical sources; the lines themselves are unchanged. --- ### A-02 · Sticky / hold-alternative for Thrust — **Class I**, default off **Problem.** Thrust is the most-held control in the game. Sustained holding is painful or impossible for players with reduced grip strength, tremor, or a single-switch input device. **Behavior.** Three modes, selectable: | Mode | Semantics | |---|---| | `hold` (default) | The line follows the key. Cabinet behavior. | | `toggle` | A press flips the line. A second press flips it back. | | `latch` | A press sets the line; **any other gameplay input clears it**. Thrust ends when you do something else, which is how thrust is used in practice. | **Why Class I and not Class M.** The thrust line is still an arbitrary boolean over time. The converter is a bounded state machine driven entirely by user edges with **no timer-driven transitions**. The cabinet's thrust button could have been held down with a matchstick and the game would behave identically. The *difficulty* of holding a button is not a Defender mechanic; the *consequence* of the line being high is, and that is untouched. **Interaction with the feedback layer.** In `toggle` and `latch`, the input overlay and the Control Trainer show thrust as a **latched** state (a filled indicator rather than a pressed one), so the player can always see the line's true value. This is required — a latched control the player cannot see is a new failure mode. **Edge cases.** On window blur, latched thrust is **cleared** along with every other held input (`SHELL_AND_LIFECYCLE_SPEC.md` §5.2) — a latched line surviving a blur means returning to a ship at terminal velocity. On death, the latch is cleared. On gamepad disconnect, cleared. --- ### A-02b · Sticky Up / Down — **Class I**, default off Same three modes for the two stick lines. Additional rule: **`toggle` and `latch` on Up and Down are mutually exclusive** — setting one clears the other, because the cabinet's 2-way stick could not report both. In `latch`, pressing the opposite direction clears the first. --- ### A-03 · Auto-fire — **Class M**, default off, **labeled** **Behavior.** While the fire binding is held, a fire edge is synthesized every **N frames**, where N is selectable: 4 (15 Hz), 6 (10 Hz), 8 (7.5 Hz). Default 6 when enabled. All three sit below the **20 Hz trigger ceiling** a human hand reaches on the cabinet button (`INPUT_MAPPING.md` §5.2), which is why the fastest option is 15 Hz and not higher — auto-fire exists to remove a motor barrier, not to exceed the hand. **Why Class M and not Class I.** The rate is timer-driven, not user-driven, and it exceeds what a hand on a cabinet button sustains. More importantly, **the laser has an on-screen cap and a cadence** (`CONTROL_SYSTEM_SPEC.md` owns the numbers), so firing rate is a real difficulty parameter. Changing it changes the game. **Label.** Settings row: *"Auto-fire — changes firing rate. Runs are marked as assisted."* Status band while active: `AUTO` in the deviation line. Score screen: *"Assisted: auto-fire (10 Hz)."* **Why offer it at all.** Rapid repeated actuation is the single most common motor barrier in arcade games and the one most likely to cause injury. A player who cannot tap cannot play Defender at all without it. Marking the run is a fair price; excluding the player is not. --- ### A-04 · One-button mode — **Class S**, default off, **labeled** **Behavior.** A single switch (or key) cycles through a radial action list and actuates. Intended for single-switch and sip-and-puff users. **Why Class S.** It cannot produce simultaneous inputs, so it cannot produce Defender's core maneuvers; conversely it makes each action free of timing pressure. The resulting game is recognisably different in both directions. There is no honest way to call the score comparable. **Design.** Dwell-to-select with a configurable dwell (0.4–2.0 s), a fixed action order (Thrust, Fire, Up, Down, Reverse, Smart Bomb, Hyperspace), a visible radial selector on the presented surface, and **the simulation is paused while the selector is open** — which is why this is Class S rather than merely M. **Label.** Persistent `1-BTN` marker; score screen states *"Not scored: one-button mode."* --- ### A-05 · Reduced flash — **Class P**, default off (on when `prefers-reduced-motion`) **What it changes.** Only the four full-screen luminance events cataloged in `HUD_RADAR_SPEC.md` §4.12. See §5 of this document for the complete analysis and the exact substitutions. **What it does not change.** Sprite-level animation, the laser's 37-color cycle, the palette cyclers, explosion particles. Those are handled by A-16 (photosensitivity-safe mode), which is stricter. **Why Class P.** `pcram` is a *shadow buffer*; the simulation writes it and the renderer reads it. Substituting what the renderer does with those values changes no simulation state. --- ### A-06 · Reduced bloom — **Class P**, default off Enhanced renderer only. Three levels: full, reduced (50 % radius, 40 % intensity), none. Classic Mode has no bloom at all, so this control is disabled and grayed in Classic with the explanation *"Classic Mode has no bloom."* **Why it matters beyond taste.** Bloom on a 1980 vector-ish palette with saturated primaries on black produces glare that measurably reduces legibility for players with cataracts, corneal scarring, or photophobia. "None" must be a real zero, not a small value. --- ### A-07 · Reduced screen shake — **Class P**, default off (on when `prefers-reduced-motion`) `shakeScale` slider 0.0–1.5, default 1.0, and a single "off" control that sets 0.0. At 0.0 **no translation is applied at all** — not a small one. Shake is Enhanced-only (`INTERACTION_FEEDBACK_SPEC.md` D-2), so in Classic this control is disabled with the explanation *"Classic Mode has no screen shake."* That sentence is also a small piece of preservation education, which is the point of the Lab. --- ### A-08 · High-contrast scanner — **Class P**, default off (on when `prefers-contrast: more`) **Problem.** The scanner is 128 × 32 pixels showing up to 100 objects in seven colors, presented at roughly 4× on a modern panel. The original palette was chosen for a phosphor CRT and several pairs are close in luminance: green `#00AE00` (L ≈ 0.30) against gray `#8989A0` (L ≈ 0.29) is a **1.02:1 luminance ratio** — effectively identical for a player with low vision. **Behavior.** Blip colors are remapped within the same hue families to maximize pairwise separation, **preserving three invariants**: 1. **Class count is unchanged** — seven visible classes, seven colors. 2. **The shimmer/steady distinction is preserved.** Pod and Mutant use animated slots; their remapped colors must still animate. This is the highest-value discrimination on the scanner (`HUD_RADAR_SPEC.md` §4.6) and losing it would be a net accessibility *loss*. 3. **Relative hue families are preserved** — the Lander stays yellow-over-green, the Bomber stays magenta — so a player switching between modes is not relearning. **The remap:** | Class | Original | High-contrast | Contrast vs black | |---|---|---|---| | Lander | `#FFAE00` / `#00AE00` | `#FFC63C` / `#37D437` | 9.1 / 8.4 | | Mutant | cycler / `#00AE00` | cycler (brightened) / `#37D437` | — / 8.4 | | Baiter | `#00AE00` | `#00FF6E` | 12.6 | | Bomber | `#FF00FF` | `#FF6BFF` | 8.9 | | Pod | cycler | cycler (brightened, wider range) | — | | Swarmer | `#FF0000` / `#FFAE00` | `#FF5B4A` / `#FFC63C` | 6.5 / 12.0 | | Humanoid | `#8989A0` | `#E8E8FF` | 16.9 | The humanoid gets the largest lift deliberately: finding humanoids is the game's central task and gray-on-black at 2 × 2 pixels is the hardest read on the display. **Why Class P.** Palette only. Every mapping is applied at present time; the framebuffer nibbles are unchanged. --- ### A-09 · Color-vision palettes — **Class P**, default off Four variants, each a full 16-entry palette replacement applied at present time: | Variant | Target | Strategy | |---|---|---| | `deuteranopia` | red-green, ~6 % of men | Red and green separated on the blue/yellow axis; the Lander's yellow-over-green becomes yellow-over-cyan. | | `protanopia` | red-green, red-dim | Reds lifted in luminance; the Swarmer's red/yellow speckle becomes violet/yellow. | | `tritanopia` | blue-yellow, rare | Blue bezel shifted toward red; yellow shifted toward white. | | `monochrome` | full achromatopsia, and CRT-monochrome authenticity | Every class mapped to a distinct **luminance** step ≥ 12 % apart, and the two cycler classes given a **distinct blink pattern** rather than a color cycle. | **The `monochrome` variant is the hard one and the most important.** With hue removed, seven classes must be distinguished by luminance and motion alone. The specification: | Class | Luminance | Pattern | |---|---|---| | Humanoid | 100 % | steady | | Player marker | 100 % | steady (already white, already a distinct shape) | | Pod | 85 % | **blink 2 on / 2 off** | | Mutant | 70 % | **blink 4 on / 4 off** | | Baiter | 55 % | steady | | Bomber | 40 % | steady | | Lander | 28 % over 55 % | two-tone, steady | | Swarmer | 28 % | steady | Blink periods are in **presented frames**, not simulation frames, and are driven by the presentation clock — so they are Class P even though the scanner itself updates at 7.5 Hz. **Rule.** No color-vision palette may reduce the number of distinguishable classes below seven. T-14 asserts this by pairwise ΔE (or Δ-luminance for `monochrome`) across all 21 class pairs. --- ### A-10 · Larger HUD — **Class P**, default 1× Scales the score, ships, smart bombs and (optionally) the scanner into an overlay drawn on the **presented** surface, at 1×, 1.5× or 2×. The framebuffer band is still drawn underneath and is simply covered. **Hard constraints, from `HUD_RADAR_SPEC.md` §9.2:** - The **6-digit score cap, the 5-ship cap and the 3-bomb cap are preserved at every scale.** Showing a seventh ship because there is now room is new information and is forbidden. - The scanner may be scaled but **its 2048-px world compression and its player-centering may not change**. A 2× scanner shows the same world at twice the size, not more of it. - At 2× on a 4:3 presentation the enlarged HUD covers roughly the top 18 % of the playfield. **The overlay is drawn with 70 % opacity over the playfield region** so nothing is fully hidden, and the player is told this in the setting's description. --- ### A-11 · Audio cue reinforcement — **Class P**, default off Adds a distinct short cue for events that are visually communicated but have **no sound at all** in the original. From the silent-event inventory (`INTERACTION_FEEDBACK_SPEC.md` §4.23): | Event | Cue | Why it earns one | |---|---|---| | Reverse fires | R-cue, 40 ms click, pitched by facing | The only control with no audio confirmation | | Hyperspace departs | H-cue, 250 ms downward glide | 15 silent black frames read as a crash | | Mutation | two-note rise, 120 ms, once per mutation | Silent, and it changes the threat class | | Planet restored | three-note rise, 300 ms | Silent, and it is a major state change | | Input refused | N-cue, 60 ms thud | Silent, and it is indistinguishable from a dropped input | | Bonus ship at the display cap | 200 ms chime layered under `RPSND` | The HUD does not change at the cap | All cues are synthesized in the same `AudioWorklet` as the game's sounds but on a **separate gain bus** with its own level control, so they can be balanced against the game or muted independently. They are **never** routed through the six-bit command bus — they are not Williams sounds and must not contaminate the audio regression corpus (`md/07` §8). **Why Class P.** Nothing in the simulation is read or written; the cue layer subscribes to the same event stream the renderer does. --- ### A-12 · Captions / event text — **Class P**, default off (one exception) Short uppercase strings in the game's own 6 × 8 font, drawn on the presented surface in a fixed two-line region **below the playfield** (in the letterbox/pillarbox margin where one exists, or in the bottom 16 presented rows otherwise). Never over the scanner. **The caption set:** ``` ABDUCTION HUMANOID CAUGHT HUMANOID LOST HUMANOID RETURNED MUTANT BAITER EXTRA SHIP + BOMB SMART BOMB NO SMART BOMBS HYPERSPACE WAVE n COMPLETE PLANET DESTROYED PLANET RESTORED SHIP LOST GAME OVER ``` Rules: - **At most 2 lines**, newest at the bottom, each held **60 frames** then removed. - **Priority-ordered**: `PLANET DESTROYED` > `SHIP LOST` > `GAME OVER` > humanoid events > everything else. A higher-priority caption displaces the older line rather than queueing behind it. - **`PLANET DESTROYED` is on by default in Enhanced Mode** (`INTERACTION_FEEDBACK_SPEC.md` §4.19) — it is a permanent, game-altering state change communicated in Classic almost entirely by a visual sequence a low-vision player cannot read. - Captions are also emitted to an **ARIA live region** with `aria-live="polite"` (`assertive` for `SHIP LOST` and `GAME OVER`), so a screen-reader user gets the same stream. See §6. - Caption text scales with A-10. --- ### A-13 · Haptic cue reinforcement — **Class P**, default on when a gamepad is present The effect table is in `INTERACTION_FEEDBACK_SPEC.md` §7.1. Accessibility-specific rules: - `hapticScale = 0` **makes no gamepad API calls at all** — some assistive controllers mis-handle zero-magnitude effects. - Haptics are **never one of the two channels required by the redundancy floor** (`INTERACTION_FEEDBACK_SPEC.md` §5.3). They are always additive. - Haptics must never affect timing (R-3). T-43 there enforces it. --- ### A-14 · Pause-friendly practice mode — **Class S**, default off, **labeled** **Behavior.** The simulation may be paused **at any moment, including mid-death and mid-hyperspace**, and single-stepped forward one frame at a time. Enemy spawns are unchanged; nothing is easier except that time stops when you need it to. **Why Class S and not M.** Unlimited think time removes the reaction-speed component entirely, which is a large fraction of what Defender measures. Being honest about that is better than pretending pausing is free. **Why offer it.** For players with slower processing speed, for players learning the scanner, and for anyone studying the reconstruction — which is, after all, the Lab's stated purpose. It is also the same machinery the Lab's frame-advance uses (`SHELL_AND_LIFECYCLE_SPEC.md` §8.2); the difference is only whether a score is claimed afterwards. **Label.** `PRACTICE` persistent marker; score screen states *"Not scored: practice mode."* **Not to be confused with ordinary pause** (A-28), which is a lifecycle feature, freezes everything, and is Class P. --- ### A-15 · Configurable hyperspace flashing — **Class P**, default `standard` Three settings: | Setting | Departure (15 frames) | Arrival (40 frames) | |---|---|---| | `standard` | screen goes black, as the original | `APVCT` appear effect at full intensity | | `soft` | black fades in over 4 frames instead of cutting | appear effect at 50 % intensity, no high-frequency component | | `minimal` | screen dims to 20 %, never fully black | a static outline that fades in over 40 frames, no flicker | **The frame budget is identical in all three (15 + 40 = 55 frames).** Only what is drawn changes. This is what makes it Class P and it is a hard constraint — a builder who "improves" the timing has changed a mechanic. --- ### A-16 · Photosensitivity-safe presentation — **Class P**, default off The strictest presentation mode. Fully specified in §5. --- ### A-17 · Slow motion — **Class S**, default off, **labeled** **Behavior.** The **host** calls `step()` at 0.75×, 0.5× or 0.25× of the 60.09615 Hz rate. **The simulation is unchanged** — every constant, every tick, every RNG call is identical; only wall-clock pacing differs. **Why Class S despite the simulation being bit-identical.** This is the case that shows why the classification is about *the run*, not about *the code*. A replay recorded at 0.25× replays identically at 1×. But the human who produced the input sequence had four times as long to produce it, and the score is therefore not a Defender score. Class S is a statement about the performance, not about determinism. **Interaction with A-14.** Both may be on. Both are S. The run is unscored either way. --- ### A-18 · Extra lives — **Class M**, default 3, **labeled if ≠ factory** The operator's `NSHIP` CMOS setting, exposed. Values 3 (factory), 5, 9. Remember that **starting smart bombs follow the ships setting** (L-024), so choosing 5 ships also gives 5 smart bombs — which the UI must state, because it is surprising and it is authentic. At the factory value this is not an assist at all and carries no label. At 5 or 9 it is Class M: *"More ships — changes difficulty. Runs are marked as assisted."* **This is the one Class M option a real operator could set on a real cabinet**, which is why it is M rather than S: high scores set on a 5-ship machine were and are a real category. --- ### A-19 · Invulnerability — **Class S**, default off, **labeled** Sets `STATUS` bit 4 (player image collision inactive) permanently. The ship still cannot pass through terrain (there is no terrain collision to begin with — L-144) and hyperspace still cannot kill (because `PLEND` is never reached). Every other mechanic runs. Offered because it is the only way some players will ever see wave 20, and because it is indispensable for studying late-wave behavior. Unscored, unmissably marked. --- ### A-20 · Hyperspace never kills — **Class M**, default off, **labeled** Suppresses the `LSEED > 192` test. **Nothing else changes** — the 15-frame black, the random destination, the zeroed velocity, the 40-frame materialise all stay. **Why M and not S.** It removes one 24.6 % risk from one rarely-used action. That is a genuine difficulty change but it is bounded and local; the rest of the game is intact. A run with it on is still recognisably a Defender run. Label: *"Hyperspace is safe — changes difficulty (normally fatal 1 time in 4). Runs are marked as assisted."* --- ### A-21 · Scanner smoothing — **Class P**, default off, **labeled anyway** Blips render at their true world position every presented frame instead of stepping on the 8-frame cadence. See `HUD_RADAR_SPEC.md` §9.3. Unavailable in Classic Mode. Labeled despite being Class P because a veteran will feel it. This is the exception that proves the labeling rule is about honesty, not about mechanics. --- ### A-22 · Larger blips — **Class P**, default off 2 × 2 framebuffer blips presented at 3 × 3 or 4 × 4 device-scaled pixels, centered on the true position. Positions and count are unchanged, so no information is added. At 4 × 4, adjacent blips can visually merge — the setting's description says so, and the default remains 2 × 2. --- ### A-23 · Input latency compensation — **not offered** **Deliberately absent.** Predicting input to reduce apparent latency means running the simulation ahead and rolling back, which breaks determinism, breaks replay, and breaks the one thing this project is for. The correct answer to latency is to have less of it: a fixed 16.640 ms step, input sampled at the top of the step, one presented frame of buffering, and no smoothing (`md/06` §10 — *"do not smooth controls at the expense of Defender feel"*). **What is offered instead:** the Lab overlay reports measured end-to-end latency (`SHELL_AND_LIFECYCLE_SPEC.md` §7.2) so a player can see and fix their own environment. --- ### A-24 · Aim assist / target lock — **not offered** There is nothing to assist. Defender's laser is a horizontal beam at the ship's altitude; "aiming" is flying to the right row. An assist would have to move the ship, which is playing the game for the player. Requests for this are answered with A-19 (invulnerability) and A-17 (slow motion), which are honest about what they do. --- ### A-25 · Shell keyboard navigation and screen-reader support — **Class P**, always on Not optional and not a setting. See §6. --- ### A-26 · Text scaling and reflow — **Class P**, follows the browser All shell chrome (the launch page, the pause menu, the settings, the score screen, the Lab overlay's text) uses relative units and reflows to 200 % zoom and to a 320 px viewport without horizontal scrolling, per WCAG 1.4.4 and 1.4.10. **The game canvas is exempt** — it is a fixed-aspect raster and scaling it is A-10's job. --- ### A-27 · Audio controls — **Class P** - **Master**, **Effects** and **Background drone** levels, 0–100 each. The drone gets its own control because it is continuous, low-frequency, and the most likely single element to cause fatigue or to mask speech for a player using a screen reader simultaneously. - **Mono downmix** — collapses the Enhanced stereo pan. Required for single-sided hearing loss, and it is also the authentic configuration (the cabinet had one speaker). - **Cue bus level** (A-11) — separate from Effects. - All levels are applied at the Web Audio graph's output gain nodes. The command stream is untouched, so the audio regression corpus is unaffected. --- ### A-28 · Pause on focus loss — **Class P**, default on Lifecycle feature; specified in `SHELL_AND_LIFECYCLE_SPEC.md` §5.2. Listed here because turning it **off** is an accessibility-relevant choice for players using external assistive software that steals focus. With it off, the game continues to run on blur but audio is still suspended and a visible "RUNNING IN BACKGROUND" marker appears on return. --- ### A-29 · Confirm-before-hyperspace — **Class I**, default off Hyperspace kills 24.6 % of the time and is bound to the pinky key. An accidental press is a lost ship. With this on, hyperspace requires **two presses within 40 frames**; the first press produces the N-cue and a 40-frame HUD prompt. **Why Class I.** It changes how the hyperspace line is produced from the user's key — a bounded state machine over user edges, no timer-driven transitions. The simulation sees an ordinary single hyperspace edge. It makes the game slightly *harder* if anything (the action takes longer), which is a useful sanity check on the classification. --- ## 5. Photosensitivity — the analysis, and what A-05 and A-16 actually do ### 5.1 The measurement We use the WCAG 2.3.1 general flash threshold as the operative rule: **no more than three flashes in any one-second period**, where a "flash" is a pair of opposing changes in relative luminance of ≥ 10 % (with the darker state below 0.80 relative luminance) occupying more than 25 % of a 10-degree visual field — approximated, as WCAG allows, as **more than 25 % of the presented area**. The reconstruction measures this directly rather than trusting the analysis: ```js // Runs on the PRESENTED surface, once per presented frame, in the feedback layer. // Never on the frame path of step(). Sampled at 1/16 resolution. luma = meanRelativeLuminance(presentedFrame, sampleGrid) areaAbove = fractionOfArea(|luma_t − luma_{t−1}| >= 0.10) if areaAbove > 0.25 and min(luma_t, luma_{t−1}) < 0.80: recordTransition() flashesPerSecond = transitionsInTrailing60Frames / 2 ``` `HUD_RADAR_SPEC.md` T-53 asserts the budget; this is the instrument it asserts against. ### 5.2 Defender's actual flash inventory, measured against that rule | Source | Flashes | Rate | Area | Verdict | |---|---|---|---|---| | **Planet destruction** — 16 full-screen color flashes, gaps of ~2 + `RMAX(i/8+1)` frames | 16 | **~15 Hz for the first half**, decelerating | 100 % | **Far over threshold. The worst event in the game by a wide margin.** | | **Smart bomb** — `COM PCRAM` ×4 at 2-frame intervals | 2 | 15 Hz burst | 100 % | 2 flashes < 3, so a single bomb passes. **Two bombs within one second = 4 flashes: over.** Reachable — the debounce allows ~20 frames between bombs. | | **Player death, step 4** — `PCRAM = $FF` for 2 frames | 1 | one-shot | 100 % | Passes alone. Combined with a smart bomb in the same second: over. | | **Player death, step 3** — ship strobes blank/draw at 2 frames through 8 colors | 8 | 30 Hz | **small** (an 8 × 6 byte-column sprite ≈ 0.2 % of area) | Under the area threshold. Passes. | | **Laser / score color cycle** — slot 1 stepped every 2 frames through `COLTAB` | continuous | 30 Hz | laser: a thin full-width beam ≈ 1–2 % of area | Under the area threshold. Passes — **but see A-16.** | | **Pod / Mutant / mine cyclers** — random `COLTAB` entry every 6 frames | continuous | 10 Hz | small | Passes. | | **Hall of Fame cursor** — 15-frame toggle | continuous | 2 Hz | tiny | Passes. | | **Player-up flash (`FLPUP`)** — 20-frame period | continuous | 3 Hz | tiny | Passes. | | **Enhanced bloom + explosion particles** | varies | varies | varies | **Must be measured, not assumed.** The instrument in §5.1 runs in Enhanced too. | **Conclusion: Defender in its original form exceeds the WCAG flash threshold in exactly two situations — planet destruction (badly) and repeated smart bombs (marginally).** Everything else is under, mostly because the flashing elements are small. That is a specific, bounded problem with a specific, bounded fix, which is much better news than "the whole game flashes". ### 5.3 A-05 Reduced Flash — the substitutions | Event | Standard | Reduced Flash | |---|---|---| | Smart bomb | 2 × full-screen white inversions, 2 frames each | **One 20 % white wash held for 8 frames.** Same duration, same beat count in the audio, one gentle rise instead of two hard inversions. | | Player death step 4 | 2 frames of `#FFFFFF` | **2 frames of 25 % white wash.** | | Planet destruction | 16 random-color full-screen flashes | **16 luminance lifts of 12 %, 2 frames each, preserving the deceleration curve exactly.** The rhythm — which is the information — is intact; the amplitude is not. | | Hyperspace | per A-15 | forced to `soft` | **The frame budgets are identical in every row.** Reduced Flash changes amplitude, never timing. ### 5.4 A-16 Photosensitivity-safe presentation — the strict mode Everything in A-05, plus: 1. **A hard governor.** The measured `flashesPerSecond` from §5.1 is clamped to **≤ 3**. If a transition would exceed the budget, the presented frame's luminance delta is scaled down until it falls below the 10 % threshold. This is a real feedback loop on the presented surface, not a static ruleset, so it also covers Enhanced bloom and particle effects that nobody analyzed. 2. **Slot-1 cycling clamped.** The laser/score color cycle is limited to a change every **8 presented frames** (7.5 Hz) instead of every 2 (30 Hz), and the `COLTAB` walk is restricted to entries within a 25 % luminance band. The color still moves — the signature is preserved — but the temporal frequency drops out of the 3–60 Hz risk band's dense region. 3. **All cyclers clamped** the same way. 4. **Explosion particle brightness capped** at 70 % and bloom forced to "none". 5. **A pre-play notice.** Before the first run in a session with A-16 **off**, the launch page shows a single dismissible line: *"This reconstruction reproduces the original's screen flashes. A photosensitivity-safe mode is available in Settings."* It is one line, it is dismissible forever, and it appears before the game starts — not as a modal over a running game. **Why Class P despite the governor.** The governor reads the presented framebuffer and writes the presented framebuffer. It never reads simulation state and never writes it. T-11 asserts replay identity with the governor active. --- ## 6. Screen readers, keyboard navigation, and the parts that are not the game The canvas is not accessible and cannot be made accessible — a 292 × 240 raster of a 1981 arcade game has no semantic structure to expose. **Everything around it must be perfect**, and the event stream must be narrated. ### 6.1 The shell | Requirement | Implementation | |---|---| | Every control reachable by keyboard | Native `