integrate Playdate target, host input, and example apps

[colmugx]

[doodlewind]

Verdict: fit to merge — this is a model target addition, and it lands with one real rendering bug fixed forward on the branch (1440211). Reviewed with the repo review workflow (skills/pocketjs-review-pr).

The finding: 1bpp polarity was inverted on device

vp_pd_render_frame wrote style 0 as raw glyph bytes and style 1 as ~glyph. Font bits are ink (vapor_esp32.c pins this: bits & 0x80>>x ? ink : paper), and Playdate framebuffer bits are white — your own device photo proves it: the bg-white text-black title bar renders white-on-black, and the dark-theme body renders dark-on-light. Every pair displayed as the exact inverse of its declared luminance orientation, of this PR's own styleTableCss playdate preview (s === 0 → ink #000 on #fff), and of what GB/NES render for the same app under the shared styles2 contract. It photographed fine because complementing every pixel preserves contrast — but it breaks the "same file, every target agrees" invariant.

Fixed to style ? glyph : ~glyph with the semantics stated in a comment, and the C harness now pins the corrected direction (frame[0] == ~0x81 for style 0). Note: a rebuilt .pdx will look color-flipped relative to your photo — dark title text on a white bar, light body text on the dark theme — which is the orientation the oracle preview and DMG render. A fresh device glance when convenient would be welcome, but the direction is proven by the photo + the ink/white bit semantics, so I didn't hold the merge on it.

Second gap: the oracle replay was claimed, not present

DESIGN.md says the Playdate Todo "is also replayed through the real Vue Vapor oracle with signed Primary-axis deltas" — no test booted todo.playdate.tsx; entry.ts hardcodes todo.tsx, so the new axisDelta hook had no listener. Rather than edit the claim down, I made it true: bootOracle takes a bundle entry, entry-playdate.ts mounts the crank variant, and three tests replay boot on the 50×30 grid, signed 45° detents with sub-detent accumulation and end-clamping, and edit-mode crank isolation. The grid-line assertions passed on first run — the app and oracle plumbing were already correct.

What was verified as-is

• Taxonomy: Playdate slots into VAPOR_TARGETS as data; the BOARDS.md addition correctly argues why a fixed SDK target is not an ESP32-style board JSON. relativeAxesUsed joins buttonsUsed as a derived demand; VT101/VT102 admission errors extend the existing VB/VS-code convention without collision.
• RelativeAxis contract: hardware-neutral signed millidegrees, apps own detents — the runtime preserves sub-millidegree remainders across frames, drains docked/lifecycle motion (no ghost events), and the C runtime harness pins all of it against a fake SDK table. The dev server's wheel adapter uses the same accumulation. This is the right shape for the future ESP32 encoder, and BOARDS.md records that follow-up honestly.
• Emitted-C changes for existing targets (condition() paren trim, app_on_axis_delta stub, VP_UNUSED_FN): behavior-neutral, proven by the GBA/GB/NES emulator per-press parity suites passing unchanged. esp32BuildId shifts for identical sources as a consequence — that is the documented contract (the id follows generated code + runtime sources; the board-half pin in boards.test.ts is untouched), so a pre-integrate Playdate target, host input, and example apps #205 MeowBit will correctly report stale on vapor:esp32:verify.
• Passing doc-rot fix: the screen.ts "NES 24x20" comment was stale; 22×18 matches VAPOR_TARGETS.nes.
• pd_api.h in the test harness is a 62-line clean-room shim (function-pointer tables + enum values, no SDK expression). No new deps; vapor/ stays outside the npm files map.
• CLAUDE.md: kept your rule, trimmed to the enforceable sentence — the design exposition lives in DESIGN.md §5.

Local gate

• bun run vapor:test: 67 pass, 0 fail (64 from the branch + 3 new oracle-replay tests; includes both C harnesses under -Wall -Wextra -Werror and the GBA/GB/NES parity replays)
• bunx tsc --noEmit -p vapor and root bunx tsc --noEmit: clean
• Root bun run test: 450 pass, 0 fail (base-relative)
• Real SDK smoke on this machine (SDK 3.x at ~/Developer/PlaydateSDK): simulator .pdx builds and validates end-to-end (51.3 KB). The device build needs the newlib-bearing ARM GNU toolchain that Panic's arm.cmake assumes — Homebrew's bare arm-none-eabi-gcc stops at stdint.h — so device evidence remains your on-hardware photo, which is exactly what runtime/playdate/README.md prescribes.