Confirmed: legacy mouse reports lose non-UTF8 coordinate bytes before reaching the PTY. One finding, independently reproduced by Claude Opus 5.5 at medium effort.
Fetched main: cb251b9abc1403aa02e21e400d0b7f396a195e3c. Actual-client proof and repeated public-owner red: 6d3f6423a341b315a091567aee5ab34a3b8b0373, the later fetched main. No terminal production change between these commits.
T3 delegates terminal behavior to its pinned Ghostty C ABI. The vendored mouse headers explicitly support normal tracking, X10 byte format, and separate UTF8 and SGR formats. This uses that existing engine contract. The ordinary raw-mode Python PTY program enables 1000 without 1005/1006/1015/1016 and reads bytes directly with os.read. XTerm specifies legacy coordinates as single-byte numeric values plus32, with UTF8 and SGR separately enabled. XTerm primary specification.
The producer requests a cursor-position response after switching modes. Receipt of that real terminal reply acknowledges mode processing before clicks. The supported configuration is the actual Linux PTY and Chromium web client, using its ordinary shell and unchanged encoding configuration. No named legacy TUI or wider platform claim is needed.
XTest sends real pointer input to Chromium. All30 captured pointerdown/up events are trusted and target CANVAS. SGR and legacy runs use the same pixel positions; SGR establishes the actual cell independently of fractional geometry. The table shows the complete legacy press packet received by the real PTY program at row4. Release packets have button byte23 in place of20.
| SGR column | Expected legacy press hex | Actual PTY press hex |
|---|---|---|
| 10 | 1b5b4d202a24 | 1b5b4d202a24 |
| 94 | 1b5b4d207e24 | 1b5b4d207e24 |
| 96 | 1b5b4d208024 | 1b5b4d20efbfbd24 |
| 99 | 1b5b4d208324 | 1b5b4d20efbfbd24 |
| 128 | 1b5b4d20a024 | 1b5b4d20efbfbd24 |
SGR reports the correct distinct coordinates in all five cases. Mode-off produces no mouse bytes at all five positions. Actual outgoing RPC frames already contain replacement characters; their UTF8 bytes match the receiver. Legacy applications expecting six-byte reports can misparse these longer packets or lose the clicked cell. Exposure is limited to legacy encoding with coordinates whose encoded bytes are not valid UTF8; SGR does not show this failure.
core.ts:1065 decodes the mouse encoder's arbitrary byte buffer as UTF8. The invalid coordinate byte becomes U+FFFD before surface.onData and RPC. The string schema and server pass the resulting string through; node-pty encodes it as UTF8 for the PTY. This conclusion follows the public-owner red and real received bytes, not the presence of a decoder alone. Simply switching to Latin1 decoding would not establish fidelity across the existing UTF8 string transport.
The existing GhosttyTerminalCore mouse test is extended in place. Columns10/95 pass;96/100/128 fail byte assertions. Original fractional targeting assertions, SGR100 and mode-off controls pass. The same extension fails at the exact runtime main SHA. No wrapper-only test or mocked wire is used as proof. The temporary tracked test extension has been reverted and retained as an uncommitted local repro patch.
Fresh read-only Opus 5.5 worktree, own VM bx_79ukz68w, reset before one serialized actual-app proof. Pinned main cb251b9abc1403aa02e21e400d0b7f396a195e3c. Existing public-owner test fails for the same three assertions; actual trusted canvas pointer inputs produce byte-for-byte identical receiver observations, and SGR/mode-off controls pass. The verifier test extension was reverted and its worktree is clean. The primary agent also read the retained verifier SHA, results and trusted-event records.
Both open and closed issue and PR searches cover mouse, terminal encoding, legacy terminal, mouse byte and mouse1006; ineligible unvouched PRs do not count as owners. No exact eligible owner was found. 37148 owns fractional coordinate mapping and bounds, which are unchanged here. This is one byte-fidelity invariant across coordinates, not separate horizontal and vertical bugs. Passed SGR all-motion/click/drag behavior is excluded.
Shift/Ctrl/Meta deliberately bypass application mouse reporting. That policy is preserved and is not a candidate. Native desktop, mobile, other operating systems, relay/tunnel deployments and named legacy TUIs were not executed. The shown proof covers the web client and real Linux server/PTY, including actual RPC. No native Ghostty reference result is claimed.
The first app harness failed before receiver readiness because typing preceded the running terminal; its sent frames contain only the command suffix. It is excluded from byte evidence. One corrected reset run waits for public readiness and completes. VM setup includes slow lazy checkout reads. No device or provider-filter retry occurred.
Hunter VM bx_5ty6jjxg and independent verifier VM bx_79ukz68w are released. All owned apps, servers, recording processes and the blank preview tab are stopped or closed. Both tracked worktrees are clean. No runtime resources remain.
No fixes, commits, pushes, issues, PRs or top-level threads. Runtime, server and capture work stayed on boat. Only the focused adjacent owner test ran on forge. Raw logs, patch and sanitized results stay in local hunt artifacts.
PostPlan uses true-black fallback colors because the active client theme could not be read.