Turn CI green, encrypt sync for real, and fix two silent no-ops

[proggeramlug]

Six commits. Two themes: CI never actually gated anything, and sync was never actually encrypted.

CI was red since May and gating nothing

The typecheck job piped tsc under set -o pipefail, so a non-zero exit killed the step before its own ~30-pattern "known noise" filter ever ran. The filter was dead code. Removed it — the step is now a plain bun run typecheck.

The "noise" it was hiding was mostly real. Typecheck is 39 → 0:

• perry/i18n had no ambient declaration at all, while 28 files import { t } from it (the IDE genuinely ships 12 locales — this import is live).
• fs / child_process declarations still described hone's old Perry fork. Added the upstream members the compat shims actually use (statSync/lstatSync/openSync + a Node-shaped Stats, spawn + ChildProcess); perry-ui was missing widgetSetTooltip.
• tests/agentic/setup/seed-files is a self-contained fixture workspace (own package.json/tsconfig, imports a .py file) copied into a temp dir by the agentic tests. It was never meant to typecheck here — excluded.

Sync advertised E2E encryption and shipped plaintext

Every crypto primitive was stubbed in both sync-host.ts and sync-guest.ts:

function ccAes256GcmEncrypt(plaintext, keyHex, nonceHex) { return plaintext; }
function ccHkdfSha256(...) { return ''; }   // -> project key stays empty

So the relay saw file deltas, AI prompts, and Claude Code transcripts in cleartext while hone.codes says "end-to-end encrypted" eight times (AUDIT-2026-07.md H1).

The stubs blamed missing Perry crypto intrinsics — true, but no longer a blocker: Perry maps node:crypto natively. Now implemented for real in a shared sync-crypto.ts (the stubs were duplicated verbatim in both files). No Rust, no FFI.

Also fixed under the same finding:

• generateProjectKey stitched two 12-byte nonces and ran them through HKDF to "reach" 32 bytes. HKDF cannot manufacture entropy it wasn't given — that was 24 bytes of randomness in a 32-byte coat. Now randomBytes(32).
• hostSecret was 'secret_' + deviceId + '_' + Date.now() — guessable from a known deviceId and an approximate launch time. Now 256 CSPRNG bits.
• Decrypt is fail-closed. It previously returned the ciphertext unchanged when the key was missing or the format was unexpected, handing unauthenticated attacker-controlled bytes back as if they were plaintext. Now returns '' on any failure and the relay handler drops the message.

Two silent no-ops that no gate could catch

Perry lowers a declared-arity mismatch to side-effect-only evaluation — it evaluates the args and returns without emitting the runtime call (perry-codegen/src/lower_call/ui_tables.rs:433-440). No compile error, no crash, no linked symbol. So a wrong hand-written .d.ts type-checks and compiles while doing nothing.

• textSetFontFamily was declared (widget, size, family); Perry's real signature is (widget, family). All 38 three-arg sites were no-ops — monoFont() has never been applied: commit-graph ASCII art, code blocks, diagnostics, chat, context chips all render proportional. Every site already set the size via textSetFontSize within ±3 lines, so dropping the redundant middle arg loses nothing. Fixed.
• scrollViewScrollTo — swept all 70 perry/ui declarations against Perry's 285-entry dispatch table (63 match; Picker wrong but zero call sites; saveFileDialog optionals hardened). This one is an upstream Perry bug and our code is the correct side: all 8 platform runtimes implement perry_ui_scrollview_scroll_to(scroll_handle, child_handle), but the table declares [Widget, F64, F64]. So scrollToBottom() is compiled away and AI chat has never auto-scrolled. Removed the invented (scrollView, x, y) overload (no runtime accepts it) and documented the bug at both the declaration and the call site so nobody "fixes" it by adding a third argument and makes it worse. Fix is one line upstream: args: &[Widget, Widget].

Also fixed a genuine crash path: findBarGetViewportStart() read viewport.visibleRange.startLine, but ViewportManager only exposes getVisibleRange() — it's wired into the find bar and would throw.

Tests

31 new. The load-bearing ones are the dullest: ciphertext is not plaintext, and the envelope's encrypted flag always describes what actually happened. Envelope construction moved to a ws-free sync-envelope.ts because sync-transport.ts imports Perry's ws extensions and cannot be imported under bun test at all — which is how the one property that mattered had no test.

Includes a regression guard pinning the exact H1 shape (with an identity "encrypt", the flag reads true while the wire carries plaintext) and a check that the PAIR_REQ| cleartext exemption is anchored, so a payload merely containing that marker can't slip past the gate unencrypted.

Build

perry compile src/app.ts now works on a clean environment — added perry.allow.nativeLibrary (Perry #497). Previously the only way through was PERRY_ALLOW_PERRY_FEATURES=1, which disables the check globally rather than approving anything. Entries match the import specifier, not the package name.

Verification

• typecheck 0 (was 39)
• 200 tests / 196 pass — the 4 failures (2 layout, 2 theme) are pre-existing and present at a14633c, before this branch
• perry compile clean on a clean env (24.4MB); app launches and renders

Notes for review

• The X25519 wire format is Perry-specific: KeyObject.export() returns an opaque surrogate string, not real SPKI/PKCS8 DER. Fine today (every Hone client is Perry-compiled), but pairing will not interop with a third-party client. Documented in sync-crypto.ts.
• Gate B is not fully closed: still needs key rotation for long-lived rooms, a device-revoke flow, and a guest/iOS on-device run.
• perry compile is still not tested in CI (if: false — no Perry toolchain on the runner). That gap is why the font bug survived: the stub and the call sites agreed with each other, just not with Perry.
• @honeide/lsp-bridge is imported by views/lsp/lsp-bridge.ts but absent from package.json dependencies, yet still resolves. Worth understanding before it bites someone on a clean checkout.