fix: remove wall-clock-timing races from 3 flaky zeph-core tests

[bug-ops]

Summary

Three tests inferred concurrency or non-blocking behavior from fixed wall-clock thresholds, which spuriously failed under CI load (confirmed on two consecutive GitHub Actions runs of the same commit on PR #6678):

• after_tool_hooks_run_concurrently_across_tier_indices (apply_tier_results_tests.rs) — replaced the elapsed-time-vs-N*delay comparison with an atomic max-in-flight high-water mark around the in-process RuntimeLayer::after_tool calls. Serial execution can never exceed 1, so max_in_flight == n is a sound structural proof.
• pre_tool_use_hooks_fire_concurrently_across_tier_indices (pre_tool_use_concurrency_tests.rs, #[cfg(unix)]) — PreToolUse hooks run as real subprocesses with no shared in-process memory, so this uses a two-phase rendezvous/witness barrier instead: every hook must individually observe all N markers present before creating a witness file, which serial dispatch can never satisfy. This replaced an earlier marker-file-poller draft that still raced a fixed sleep against subprocess-start spread (caught in review — the original CI failure showed ~980ms spread for 4x60ms hooks, exactly the kind of assumption this rewrite removes).
• dump_request_smoke_returns_promptly_and_write_still_lands -> dump_request_returns_before_blocking_write_completes (debug_dump/mod.rs) — replaced a 50ms wall-clock budget with a test-only TestWriteGate channel handshake (30s bound, so a regression to an inline write path fails with a clear panic instead of hanging the CI shard) that proves dump_request returns before its spawn_blocking write completes, with no timing race window at all.

No production behavior changed — only test assertions/instrumentation, plus a #[cfg(test)]-only field on DebugDumper.

Closes #6679

Test plan

• Targeted 10x+ repeated loop for the 3 rewritten tests, multiple rounds across the iteration cycle (final count: 0 flakes across 100+ combined runs)
• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins (15154 passed, 0 failed)
• Rustdoc gate (RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler")
• Independent adversarial review confirmed the barrier logic is sound (no false-pass path under serial dispatch, no hang path on regression) and the #[cfg(unix)] gating is complete