Fix: consolidate the three AgentLoader process-handler PRs onto one branch

[AmaadMartin]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

2. Or, if no issue exists, describe the change:

Problem: Three open PRs rewrite the same process-handler block in dev/src/utils/agent_loader.ts, so whichever lands first makes the other two unmergeable. Each one alone leaves a real defect: #511 fixes the listener leak and the silent uncaughtException exit, #703 fixes the exit status, #653 fixes the temp-directory leak. On main the constructor installs five listeners that nothing removes, Ctrl-C on adk web exits 0, and every run leaves an adk_agent_loader-* directory in the OS temp directory.

Solution: This branch carries all three fixes on one trunk. #511 is the trunk because opt-in registration is the only one of the three that changes the shape of the block, so the other two reduce to a small delta on top of it. Commits 1-5 are #511 verbatim, rebased on main with no conflicts. Commit 6 folds in #703's exit status and commit 7 folds in #653's synchronous cleanup.

Collision check: gh pr list plus a diff grep for process.on|process.exit|installProcessHandlers|disposeAllSync|uncaughtException over every open PR that touches agent_loader.ts (#830, #814, #793, #755, #731, #674, #662, #854). Only #511, #703 and #653 touch this block, so nothing else collides.

Behaviour changes (all intended, all argued in the source PRs):

1. A crash now exits 1 with a stack, and Ctrl-C exits 130. Both were 0.
2. Handler registration is opt-in. All five construction sites in this repo are updated; AgentLoader is not exported from core, so the blast radius is the dev package.
3. AdkApiServer.stop() disposes the loader it built itself. A caller who injected a loader is unaffected.

Conflict resolutions worth review:

• One onSignal closure reads the signal name Node passes it, instead of Fix: exit non-zero when an AgentLoader process crashes or is interrupted #703's array of three [signal, closure] tuples. This keeps Fix: make AgentLoader process exit/signal handlers opt-in and removable #511's loop shape.
• disposeAllSync() also calls removeProcessHandlers, which Fix: clean up compiled-agent temp directories synchronously on process exit #653 did not. Removing a listener while 'exit' is being emitted is allowed, and the symmetry keeps disposeAllSync() usable for ordinary teardown.
• The three PRs each added a listener-tracking test harness. I kept Fix: make AgentLoader process exit/signal handlers opt-in and removable #511's (listenerAddedBy, counts, afterInstall) and promoted it to module scope. Fix: clean up compiled-agent temp directories synchronously on process exit #653's trackLoaderProcessListeners().restore() only existed because the old constructor gave no way to remove listeners; disposeAll() does that now.

Known residual leak, not fixed here: adk web can still leave one scratch directory when the debug UI compiles an agent, the server disposes that borrowed AgentFile through await using, and a later request recompiles it. disposeSync() skips a file already marked disposed. The root cause is the borrowed-handle lifetime, which #830 fixes.

Testing Plan

Please describe the tests that you ran to verify your changes. This is required for all PRs that are not small documentation or typo fixes.

Unit Tests:
[x] I have added or updated unit tests for my change.
[x] All unit tests pass locally.

npx vitest run --project unit:dev dev/test/utils/agent_loader_test.ts     # 54 passed
npx vitest run --project unit:dev dev/test/server/adk_api_server_test.ts  # 58 passed
npx vitest run --project unit:dev                     # 276 passed, 1 failed
npm run build && npx eslint <the four changed files> && npx prettier "dev/**/*.ts" --check

The one unit:dev failure is pre-existing and unrelated: dev/test/cli/cli_create_test.ts > should handle Vertex AI selection with gcloud defaults reads the developer's real gcloud config. It fails identically with dev/ checked out from main.

npx tsc --noEmit reports no error in any file this PR touches. The errors it does report are pre-existing, in core/test/ and tests/integration/.

Mutation proofs. Every new test was run once against the unfixed line.

Mutation	Test that failed	Message
process.exit(128 + os.constants.signals[signal]) -> process.exit()	exits SIGINT/SIGUSR1/SIGUSR2 with 128 plus the signal number	expected "spy" to be called with arguments: [ 130 ]
removeListener -> removeAllListeners	keeps another loader listeners when disposeAll runs twice	expected [ 0, 0, 0, 0, 1 ] to deeply equal [ 1, 1, 1, 1, 1 ]
onExit -> () => void this.disposeAll()	cleans up synchronously from the exit listener, the exit listener disposes cached agents	expected true to be false, expected "disposeAllSync" to be called at least once
drop removeProcessHandlers from disposeAllSync()	removes installed listeners on disposeAllSync	expected [ 1, 1, 1, 1, 1 ] to deeply equal [ 0, 0, 0, 0, 1 ]
drop fs.rmSync from disposeSync()	5 tests, incl. removes the compiled output directory synchronously on disposeSync	expected true to be false
ownsAgentLoader = false	should dispose the agent loader it built itself when stopped	expected [ 'agent-0' ] to deeply equal []
stop() finally -> straight-line	should release its own agent loader even when closing the server fails	expected "disposeAll" to be called 2 times, but got 1 times

Existing tests this PR edits. Both edits sit in the commit that changes the behaviour they pin.

• the exit listener disposes cached agents now spies on disposeAllSync, because that is what the listener calls. The behaviour it pins is unchanged.
• The server suite's file_utils stub moves from getTempDir to createTempDir. main removed getTempDir, so the old stub was inert and the ownership test failed with expected [] to have a length of 1.

Tests from the superseded PRs that are not carried over. Each is a duplicate; the surviving test is named.

• Fix: exit non-zero when an AgentLoader process crashes or is interrupted #703 does not install an uncaughtException listener -> never installs an uncaughtException listener.
• Fix: exit non-zero when an AgentLoader process crashes or is interrupted #703 installs one exit listener and one per termination signal -> installs exit and termination signal listeners on demand.
• Fix: exit non-zero when an AgentLoader process crashes or is interrupted #703 removes every listener it installed on disposeAll -> removes installed listeners on disposeAll.
• Fix: exit non-zero when an AgentLoader process crashes or is interrupted #703 disposes the loader from the exit listener -> the exit listener disposes cached agents.
• Fix: clean up compiled-agent temp directories synchronously on process exit #653 exits the process from the SIGINT handler, which emits the exit event -> exits SIGINT with 128 plus the signal number and cleans up synchronously from the exit listener.

Manual End-to-End (E2E) Tests:

npm run build
mkdir /tmp/agents && cat > /tmp/agents/temp_agent.js <<'EOF'
export const rootAgent = {name: 'tempAgent', [Symbol.for('google.adk.baseAgent')]: true};
EOF
npx adk web /tmp/agents --port 8080     # then GET /list-apps, then Ctrl-C
echo $?                                 # 130
ls -d $TMPDIR/adk_agent_loader-*        # the directory this run created is gone

Measured on this branch, for both adk web and adk api_server: exit 130, scratch directory removed. Measured on main with the same script: exit 0, scratch directory left behind.

tests/integration/ needs no edit. The integration project does not run in my sandbox: its fixtures run npm install, which cannot reach the registry here. It fails the same way on main.

Checklist

[x] I have read the CONTRIBUTING.md document.
[x] I have performed a self-review of my own code.
[x] I have commented my code, particularly in hard-to-understand areas.
[x] I have added tests that prove my fix is effective or that my feature works.
[x] New and existing unit tests pass locally with my changes.