[ThomasK33]

Why

The user wants four small leftovers from the mouse-encoding PR (#12) fixed and landed on main. This PR fixes one of them.

normalizeMouseOptions and normalizeMouseEvent in src/index.ts read some caller properties more than once: once to validate, again to copy. A getter can pass the TypeScript checks with one value and hand the addon another. The addon checks again, so memory stays safe. But the TypeScript layer then validates a value it does not forward.

On main with a ReleaseSafe addon, x is read twice and geometry.screenWidth three times. A getter that returns 3 and then 1e30 for x throws a RangeError for a value the caller's first read never produced. A screenWidth getter that returns 80 and then 0 passes the TypeScript check and fails in the addon.

What changed

• Both normalizers destructure each caller field into a local at the top, including options.geometry, every geometry field, anyButtonPressed, trackLastCell, event.action, button, x, y, modifiers and its shift, ctrl, alt. They validate and forward only those locals.
• Error types, messages and check order are unchanged.
• A new test in the native suite feeds getters that change after the first read and expects the bytes for the first value.
• src/README.md now states that each caller field is read once.

Scope

Only the double read in the two TypeScript normalizers. The other three #12 leftovers are separate PRs. native/ is unchanged.

Blast Radius

Callers with plain data objects see no change. Callers whose getters return different values on each read now get the first value checked and encoded. A getter that throws now throws when the normalizer destructures it, before any later field check. That changes which error such a caller sees, not the message of any check.

Verification

• Test first, on commit 5d2aa49 (test only): npx vitest run test/terminal.test.ts -t "getters change" failed with RangeError: mouse x is outside Ghostty's cell or pixel coordinate range at normalizeMouseEvent src/index.ts:130.
• Same command on 01f4576 (fix): 1 passed.
• Repro script after the fix (ReleaseSafe addon, node repro.mjs). The R1 lines belong to another leftover and are the same as on main:

  R2 x getter (3, then 1e30): bytes 1b 5b 3c 30 3b 34 3b 34 4d
  R2 x reads: 1
  R2 geometry.screenWidth getter (80, then 0): bytes 1b 5b 3c 30 3b 34 3b 34 4d
  R2 screenWidth reads: 1
  

  On main the same script printed x reads: 2 and screenWidth reads: 3 and threw both errors.
• npm ci && build:native (ReleaseFast) && npm run build && npm run verify: exit 0, 68 tests passed.
• Same with ReleaseSafe: exit 0, 68 tests passed.
• UBSan (rm -rf dist, CXXFLAGS=-fsanitize=undefined,float-cast-overflow -fno-sanitize-recover=all, npm test && npm run smoke): exit 0, 68 tests passed.
• .agents/skills/verify-native-lifecycle/run.sh: OK: every result matches EXPECT.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: medium

[ThomasK33]

Verdict: PASS+NOTES
Head: 01f4576

Parent: bf414593fa85a87f417f6efa76be9969410ffaee (merge base, equal to main). Every build and test ran on Node v26.12.1 with unset Path. npm exec -c 'node -p process.version' and a one-line test under npx vitest run both printed v26.12.1.

• Scenario 1, repro. I ran node repro.mjs from each checkout root against dist/ with the ReleaseFast addon.
  • Parent: the x getter (3, then 1e30) threw RangeError: mouse x is outside Ghostty's cell or pixel coordinate range with x reads: 2. The geometry.screenWidth getter (80, then 0) threw the addon's RangeError: screenWidth must be a positive 32-bit integer with screenWidth reads: 3.
  • Head: both R2 cases print bytes 1b 5b 3c 30 3b 34 3b 34 4d, with x reads: 1 and screenWidth reads: 1. The head prints the same with the ReleaseSafe addon.
  • The R1 lines are identical at parent and head.
• Scenario 2, plain objects. My script calls encodeMouse in SGR (1000+1006), any-motion (1003) and SGR-pixels (1016) modes. It covers press, release, motion with and without a button, anyButtonPressed, trackLastCell, modifiers, padding and pixel geometry, plus 54 invalid inputs. 15 of the invalid inputs set two bad fields at once, which pins the check order. It prints 100 lines: 46 byte results and 54 ErrorType: message results. diff of the parent output against the head output is empty.
• Read counts. I wrapped all 19 caller fields in counting getters. These are the 5 event fields, the 3 modifier booleans, options.geometry, anyButtonPressed, trackLastCell and the 8 geometry fields. At head each field is read exactly once, and the bytes match the parent's. At the parent 10 fields are read more than once (event.button and event.modifiers 5 times, action, x, y and the four geometry sizes 3 times, options.geometry twice).
• Gate at head.
  • ReleaseFast: npm ci, build:native, npm run build, npm run verify. Exit 0, 68 tests passed in 5 files. The parent has 67.
  • ReleaseSafe: same steps, exit 0, 68 passed.
  • UBSan in the CI order (rm -rf dist, -fsanitize=undefined,float-cast-overflow -fno-sanitize-recover=all, npm test && npm run smoke): exit 0, 68 passed, smoke printed its snapshot, and the logs have 0 runtime error lines. nm -u shows __ubsan references in addon.o, terminal.o and libghostty_vt_shim.o.
• Revert test. At head I ran git checkout <parent> -- src/index.ts, then npx vitest run test/terminal.test.ts -t "getters change". The new test failed with RangeError: mouse x is outside Ghostty's cell or pixel coordinate range at normalizeMouseEvent src/index.ts:130. With the head file restored, it passed.
• Diff. The PR changes the two normalizers in src/index.ts, adds one test and changes one README sentence. native/ is unchanged. The new README sentence matches the read counts above. Renaming modifiers to normalizedModifiers and dropping the ! assertions both follow from the destructuring. Read order changes only for getters, as the Blast Radius section says.

Note (non-blocking):

• The new test covers 2 of the 19 fields (x and geometry.screenWidth). If a later change reads another field twice, such as y or cellWidth, the test still passes. A test that wraps every field in a counting getter and expects one read each would guard the whole README sentence.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: xhigh