fix: restore short runtime observation containment

[schickling-assistant]

Problem

PR #78 turned a consumer-side timeout increase into a fleet-scalability invariant. Its 45-row test did not exercise fleet observation: the fake pty slept for three seconds before printing a prebuilt JSON file. The resulting 10-second deadline delayed fail-closed diagnosis without fixing the observation boundary.

Goal

Keep a short outer bound around a wedged runtime, remove the synthetic latency contract, and leave fleet scalability with the runtime producer and pure exact-ID observation architecture.

Decisions

• Restore the two-second outer containment bound instead of calibrating a larger global-census timeout.
• Remove the fixed-sleep fleet test and its invariant; PR test: gate st2 against bounded PTY fleet observation #84 supplies the real packaged 0/75/100/500-session contract.
• Preserve the PATH-resolved sleep in the hung-process fixture because that portability repair is independent of the timeout policy.
• Do not add a compatibility facade or implement an st2-only "exact" API over global enumeration. R19's pure exact-ID observation boundary remains separate cross-repository work.

Verification

• nix develop -c cargo test --test doctor doctor_bounds_a_hung_pty_probe_and_reports_the_runtime_error -- --exact — 1 passed; the proof completed in 2.02s and observed timed out after 2.0s.
• nix develop -c cargo test --test up_once_exit — 2 passed.
• nix develop -c cargo clippy --all-features --lib --bin st2 --test doctor --test up_once_exit -- -D warnings — passed.
• nix flake check --print-build-logs — all six x86_64-linux checks passed, including packaged tests and invariant-reference resolution.
• git diff --check — passed.
• nix develop -c cargo test --all-targets --all-features — affected suites passed; the run later hit the pre-existing up_materialize_only_writes_the_overlay_without_needing_pty PATH fixture failure. The exact failure reproduces unchanged on origin/main.
• Repository-wide cargo fmt --all -- --check remains red on unrelated files already unformatted on main; the changed diff is whitespace-clean.

Complexity

No new complexity. This removes 71 net lines and restores the timeout as a defensive boundary.

Concerns

PR #84 also touches tests/doctor.rs, so merge order may require a small textual rebase. The semantic contracts are complementary: #84 proves real fleet behavior; this PR removes the synthetic latency invariant.

Friction & bottlenecks

• Friction: the broad Cargo test's PATH-isolation fixture removes git, producing a baseline failure unrelated to this change.
• Friction: the repository-wide rustfmt gate is not clean on current main.
• No new measured bottleneck was introduced.

Follow-ups

• Merge the real packaged fleet contract in test: gate st2 against bounded PTY fleet observation #84.
• Implement R19 through a pure exact-ID observation API; keep whole-registry snapshots at explicit audit/eval-cleanup seams.

References

• Supersedes fix: tolerate bounded slow PTY fleet census #78.
• Refs test: gate st2 against bounded PTY fleet observation #84.
• Refs R19 in docs/vrs/requirements.md and docs/vrs/spec.md.