test(e2e): harness answers DSR-6n cursor queries

[fentas]

The gap

The e2e harness is a screen scraper, not a terminal emulator: it spawns atty on a PTY and reads everything atty prints into a VT grid, but it never writes anything back in response. A real terminal answers ESC[6n with ESC[<row>;<col>R — ours never did.

So under test, no cursor replies exist at all, and every DSR round-trip is unobservable: atty's own prompt re-anchoring, and any foreground child (atuin's Ctrl+R) that queries the cursor. That's a structural blind spot over src/cursor_dsr.zig + the proxy's stdin interception — the subsystem CLAUDE.md flags as delicate.

What this adds

• dsr_reply on|off DSL verb (also settable pre-spawn, and togglable mid-scenario).
• A responder in Session.pumpMs: after grid.feed, every ESC[6n in the chunk is answered with the grid's current cursor as ESC[<row>;<col>R, 1-based.
• The dsr_child_reply scenario.

Opt-in by design. Default is off, so no existing golden changes — verified by a full-suite run: 36 pass, 0 fail, 0 updated (both known flakes included).

Two implementation notes:

• The responder writes straight to the master, not via writeInput — the latter drains through pumpMs when the buffer is full, which would re-enter the pump it's called from (and formed a circular inferred error set).
• The reply is computed after grid.feed, so the position reflects the chunk just drawn.

Honest scope of the bundled scenario

dsr_child_reply is a characterisation test of the contract "a foreground child receives its cursor reply." It passes on today's atty and is not a regression test for the atuin hang — that failure is still unreproduced (draft #553).

Its real value so far was the opposite of the usual: this harness capability is what let me disprove my own fix in #553. Three separate repro attempts, built on this responder, all passed against unfixed master — which is why #553 is held as a draft instead of merged. A test rig that can falsify a proposed fix is worth having even when it isn't yet catching the bug.

zig fmt, zig build, zig build test, and the full e2e suite are green.

[Copilot]

Pull request overview

Adds opt-in terminal-like DSR-6n cursor query answering to the Zig e2e harness so scenarios can observe cursor round-trips (needed for validating atty’s DSR interception behavior and foreground-child interactions like atuin’s Ctrl+R).

Changes:

• Extend the e2e DSL with dsr_reply on|off and thread the flag through the runner into the session.
• Implement a DSR responder in the harness pump loop that replies to ESC[6n with ESC[<row>;<col>R (1-based) using the VT grid cursor.
• Add a new dsr_child_reply characterization scenario + per-scenario config to exercise the contract.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/e2e/dsr_child_reply/scenario.e2e	New characterization scenario that asserts a foreground child receives a DSR cursor reply.
tests/e2e/dsr_child_reply/golden/env.toml	Recorded environment snapshot for the new scenario.
tests/e2e/dsr_child_reply/config.zig	Scenario-specific config enabling statusbar to ensure atty issues its own DSR queries.
src/test/e2e/runner.zig	Plumbs dsr_reply from DSL into session (pre-spawn and mid-run toggle).
src/test/e2e/harness.zig	Adds Session.dsr_reply and emits DSR replies during pumpMs.
src/test/e2e/dsl.zig	Adds dsr_reply directive parsing and command kind.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

src/test/e2e/harness.zig:143

• answerDsr writes synthetic cursor replies directly to the PTY master but does not record them in cast.json as input events. That makes the asciinema recording incomplete and can hinder replay/debugging of dsr_reply on scenarios.

            var buf: [32]u8 = undefined;
            const reply = std.fmt.bufPrint(&buf, "\x1b[{d};{d}R", .{
                self.grid.cur_row + 1,
                self.grid.cur_col + 1,
            }) catch continue;

src/test/e2e/runner.zig:375

• When dsr_reply is toggled mid-run, session.dsr_match is not reset. If the toggle happens while a partial \x1b[6n match is in progress, re-enabling can mis-detect or miss the next query. Reset the incremental matcher whenever the feature is toggled.

            // Also honoured pre-spawn (applied at session creation); allowed
            // here so a scenario can toggle the terminal's DSR answering
            // mid-run.
            .dsr_reply => session.dsr_reply = c.int_arg == 1,
            .type_str => try session.typeWith(c.str_arg, @enumFromInt(c.int_arg)),

[fentas]

Review loop complete — 2 rounds (Copilot + subagent in parallel).

Round 1 — 6 findings addressed:

• Subagent (fix-and-ship): a ESC[6n split across reads was never answered (per-chunk substring search) → incremental matcher carrying state across reads. The assertion accepted any R-terminated blob → the child now validates the reply shape. dsr_reply moved into SpawnOpts so it is live before the first pump. DSL grammar header + parse tests added. Scenario reframed around the contract it guards.
• Copilot: the reply write dropped on EINTR/EAGAIN (silently stranding whoever queried → intermittent timeout) → bounded retry. Runner pass 1 applied .dsr_reply while scanning the whole script, so a post-spawn directive also set the initial state → now gated on !spawn_seen.

Round 2 — 4 findings addressed:

• Subagent (ship): golden/env.toml still carried a stale ATTY_TRACE = "cursor" from debugging → regenerated. The scenario comment overclaimed (a leaked atty CPR is itself CPR-shaped, so the shape check separates a cursor report from anything else, not whose reply it is) → corrected. Documented the end-of-chunk cursor limitation. A dsr_reply toggle now clears partial-match state.
• Copilot: no findings.

Verification: full e2e suite 36 pass / 0 fail / 0 updated (twice — before and after the matcher change), and the scenario was mutation-checked to FAIL with dsr_reply off by me and independently by the round-2 reviewer.

Subagent verdicts: fix-and-ship → ship. All threads replied + resolved.