[novoj]

Fixes what the 2.11.0 test-flight run and the PR #46 review found.

Closes #48
Closes #49
Closes #50
Closes #51
Part of #47: § 1–3 are fixed; § 4 has three of its live cases (see "Not in this PR").

Server (#48, #50, #47)

• Probes wait for their class (B1). jdwp_set_fast_logpoint / jdwp_set_fast_breakpoint on a class that is not loaded yet register a waiting probe instead of refusing. It arms on the class's ClassPrepareEvent, keeps its ID, and shows as WAITING_FOR_CLASS in the overview.
• Hits are read at the target's exit (B2). While a probe is armed, a SUSPEND_ALL VMDeathRequest holds the target at VM death until each ring buffer is drained (bounded at 5 s). This also works under System.exit, which is how surefire's forked JVM ends. The [VM_DEATH] reply lists what the run recorded and each field watchpoint's hit count, with a note when a watch saw nothing.
• Reset counts manual holds (B3). jdwp_reset reports the jdwp_suspend_thread holds it released. jdwp_reset and jdwp_disconnect say when the target has already exited.
• Stops say what stopped.
  • Exception stops give the type and message.
  • Field-breakpoint stops give the access site and the old and new values.
  • Each fast-breakpoint stop keeps its own reason (Follow-ups from the PR #46 code review: per-stop reasons, per-invocation watchdog timing, late ring writes, race tests #47 § 1).
  • A stop names the logpoint hits recorded since the last jdwp_get_events.
• jdwp_set_exception_breakpoint takes suspendPolicy (thread | all).
• Evaluation.
  • Checked exceptions are allowed (throws Throwable).
  • Unqualified calls to the frame's methods work.
  • The thread defaults to the current stop.
  • Compile errors are listed once, with the offending code.
• Frames and locks.
  • JDK frames inside the test harness are collapsed.
  • A monitor-blocked frame shows the synchronized line: HotSpot reports the line after it.
  • jdwp_dump_locks lists Object.wait() threads separately from the lock graph.
• Watchdog (Follow-ups from the PR #46 code review: per-stop reasons, per-invocation watchdog timing, late ring writes, race tests #47 § 2). It times each invocation, not the whole lease. Two blocked samples must name the same monitor.
• Late ring writes (Follow-ups from the PR #46 code review: per-stop reasons, per-invocation watchdog timing, late ring writes, race tests #47 § 3). A probe write counts its writers, and a clear waits for in-flight writes before the last drain.

Plugin install (#51)

hooks/ensure-server-jar.sh builds the server JAR at session start. It tries ./mvnw, then mvn -o, then mvn, and says when all three fail. The jdwp-inspector entry moved into .claude-plugin/plugin.json, and the root .mcp.json is gone. After a hook-triggered build, the server needs a reconnect in /mcp, and the hook says so.

Test flights (#49)

• Flights fix: make breakpoint registration passive by default (forceLoad opt-in) #5, Expression evaluator is single-statement only — add { block } mode #8 and fix: emit eval wrapper into target's package when this is non-public (#7) #10 match what the tools do.
• fix: make breakpoint registration passive by default (forceLoad opt-in) #5 teaches the field-watchpoint blind spot: a reflective Field.set trips no watchpoint. Its par is now 3.
• Expression evaluator is single-statement only — add { block } mode #8 is par 4, and fix: emit eval wrapper into target's package when this is non-public (#7) #10 has a 4-call par path.
• The scoring rule is written into both READMEs.
• No flight gives hints any more:
  • test comments, which players may read, no longer name the root cause;
  • the class Javadocs no longer state the bug;
  • the README Hint: lines are gone.
• The sandbox zip requires plugin 2.11.1 or newer, because fix: emit eval wrapper into target's package when this is non-public (#7) #10's par path depends on B1 and B2.

Tests

• Unit: 1379 tests, 0 failures. The 43 skipped are the opt-in live tests.
• The listener, session and watchdog tests pass on a single core (taskset -c 0).
• Live suites against forked JVMs pass, including:
  • the new FastProbeLifecycleLiveTest: waiting probes, and hits read at exit after main returns and under System.exit; it passed 6 runs in a row;
  • three new SuspensionOwnershipLiveTest cases: reset during an evaluation, co-delivered arming, and disconnect during an evaluation that never returns.

Not in this PR

• Follow-ups from the PR #46 code review: per-stop reasons, per-invocation watchdog timing, late ring writes, race tests #47 § 4 deterministic barrier tests: none of the four are written yet.
• Follow-ups from the PR #46 code review: per-stop reasons, per-invocation watchdog timing, late ring writes, race tests #47 § 4 live cases still missing: a stop-all stop handed to a running evaluation, and a cancelled thread-policy set followed by a new stop of the same thread.
• Exception stops with suspendPolicy=all: no end-to-end test yet. They take the same route through SuspensionSession as breakpoint stops.
• Flight fix: emit eval wrapper into target's package when this is non-public (#7) #10 on the new build: not flown yet.