fix: surface silent adapter launch failures as session errors

[edwin-zvs]

Fixes #1208.

Problem

When a wrapper adapter's harness child exits immediately with a fatal configuration error (e.g. grok rejecting --allow "MultiEdit(...)" with unknown tool prefix: MultiEdit), the error only reached the adapter's stderr log via emit.log. The session showed awaiting input with an empty transcript — indistinguishable from a hung agent until someone grepped the daemon log.

Change

• adapter-common: new StderrTail (bounded, last 20 lines) + spawn_stderr_tail (same as spawn_stderr_log but retains the tail), and emit_launch_failure_if_silent — after a completed turn, if the child exited non-zero and no session event was emitted since spawn, emit a SessionEvent::Error carrying the exit status and captured stderr (or a pointer to the daemon log when stderr was empty).
• protocol: EventEmitter now counts emitted session events (events_emitted(), shared across clones) so adapters can tell a failed turn that produced output apart from one that died silently at launch. Turns that produced real output and then failed are left alone — the transcript already shows what happened.
• Wired into the headless turn loops of claude, codex, grok, pi, hermes, muse. Only TurnOutcome::Completed turns are checked — interrupted/stopped turns are killed by us and would misreport the kill as a launch failure. Codex keeps its existing ERROR:-line detection as the primary path; this is the fallback. Hermes keeps surfacing stdout first and falls back to the stderr tail when stdout was empty.

Regression test

crates/adapter-common launch_failure_tests: spawns a real child that prints the exact failure from #1204 to stderr and exits 2, drives it through drive_turn, and asserts exactly one SessionEvent::Error containing the stderr is emitted. Negative cases: clean exit, and non-zero exit after real output (no synthetic error). Plus tail-bounding and empty-stderr message tests.

Not in scope

• Interactive (run_pty) launch failures — the PTY path renders the child's dying words in the terminal, and the issue's repro is headless.
• Per-adapter tool-prefix validation/feature detection (issue's suggestion 3) — worth its own issue if wanted.