frontier --watch: rerun a batch side by side, specs sharing a file stem on one worker (hot reload H11) #8852

Part of #8737 (hot reload); a follow-up to H5 (#8751) and #8825.

Problem

• Cost: a save to a spec that others import changes the closure digest of every importer, and frontier --watch reruns up to WATCH_BATCH (16) of them one after another. The last verdict therefore arrives after the sum of their test times. specs/verified/reuse.t27 alone has 8 direct importers.
• Measured (2026-10-11, load ~7): 8 warm t27c test-report runs took 17.8 s one at a time and 8.3 s side by side.

Constraint

• test_report::run names its scratch directory t27c-test-report-<file stem>-<pid>. Two threads of one process on two specs with the same stem would delete each other's binaries; pattern.t27 lives in three directories.
• Changing test_report.rs would move every seal's test_runner hash (the owner's H2 decision), so the fix stays outside it.

Change

• The schedule: the batch is grouped by file stem. Groups run side by side through par_in_order (suite_jobs() workers, 8 MiB stacks); each group's specs run one after another on its worker.
• Output: verdicts print in path order, as before.
• watch.t27's comment describes the schedule.

Acceptance criteria

• After a save to specs/verified/reuse.t27, the verdict lines are identical, in the same order, with T27_SUITE_JOBS=1 and with the default.
• The last verdict arrives sooner with the default.
• Hand-written foreign lines do not grow.

Boundary

specs/verified/watch.t27

User Scenarios

• Given a spec specs/verified/reuse.t27 with 8 direct importers, When the file is saved while frontier --watch is running, Then the 8 importers are grouped by file stem and each group runs on a separate worker in parallel, with specs within a group running sequentially.
• Given T27_SUITE_JOBS=1 and the default (8 workers), When the same save triggers a watch batch, Then the verdict lines printed are identical and in the same path order.
• Given the parallel schedule, When the batch completes, Then the last verdict arrives sooner than the sequential baseline.

Requirements

• FR-001: The watch batch scheduler MUST group specs by file stem so that specs sharing a stem run sequentially on the same worker.
• FR-002: The scheduler MUST execute groups side by side using par_in_order with suite_jobs() workers and 8 MiB stacks.
• FR-003: Verdicts MUST be printed in path order, identical to the sequential output.
• FR-004: The change MUST NOT modify test_report.rs or any seal's test_runner hash.
• FR-005: The schedule logic MUST be described in specs/verified/watch.t27's comment.