fix(#465): java attach actually pauses the JVM; get_stack_trace gains threadId

[debugmcpdev]

Closes #465.

Root cause

Not a frameless-thread anchor after all: the JDI bridge never suspends the VM on attach (its attach handler only suspends for stopOnEntry in the DAP args, which the transform doesn't send), and java's policy lacked pauseAfterAttach — the only mainstream policy without it. So the session reported paused while the JVM kept running, and every thread refused stackTrace ("not suspended") — which is why the existing frameless-thread scan found nothing anywhere, and why continue → pause (a real vm.suspend()) fixed it.

Fix (issue asks 1–4)

1. JavaAdapterPolicy.getAttachBehavior → pauseAfterAttach: true, pauseAllThreads: true (new knob): the post-attach pause is sent with threadId 0, so the bridge suspends the whole VM and anchors its stopped event via its own pickReportableThread hardening ([BUG] Java attach pause_execution doesn't actually suspend the target #352) — a thread that can actually report frames.
2. get_stack_trace accepts threadId (ids from list_threads); when that thread reports frames it is adopted as the anchor for follow-up scopes/locals/evaluate. The escape hatch the old note pointed at now exists. (No silent re-anchor scan on explicit requests — an empty answer for the asked-about thread stays honest.)
3. The note names the real recovery — threadId targeting or continue_execution + pause_execution — instead of advice that couldn't be followed.
4. get_local_variables no longer contradicts session state: paused-with-frameless-anchor now says exactly that, with the same recovery pointers.

Verification

• Live repro (java -agentlib:jdwp=...,suspend=n, attach_to_process): get_stack_trace returns PauseTest.main and get_local_variables returns counter on the first try (previously: empty frames, unactionable note, "may not be paused"). threadId: 1 targeting verified too.
• New unit tests: java sends pause-all (threadId: 0) post-attach; policy pin for the attach behavior; the policies-without-behavior pin moved from java to go (java joining was the point).
• Full unit suites: 3207 passed.

Note: the issue's C/C++ observation (pause anchoring to an arbitrary thread, e.g. LLDB's injected DbgBreakPoint thread) now has a complete discoverable path: list_threads → get_stack_trace {threadId} → locals anchored there. Composes with #482's frame walk-down.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 68.75000% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/server.ts	58.33%	5 Missing ⚠️

📢 Thoughts on this report? Let us know!