[v8-bridge] surface unhandled promise rejections via cs_v8_last_error

[cs01]

Before

Unhandled promise rejections in @chadscript: interpret pragma code vanished silently. A script like new Promise((_, rej) => rej("boom")) with no .catch() would run, SpinEventLoop would drain, and cs_v8_last_error stayed empty — the caller treated the script as successful.

After

V8's PromiseRejectCallback is wired on isolate init. Unhandled rejections (event kPromiseRejectWithNoHandler — specifically ignoring the kPromiseHandlerAddedAfterReject false-positive) are captured into a thread-local. After SpinEventLoop returns, both run_script and cs_v8_eval_script_node propagate the captured rejection into t_last_error and fail the call so cs_v8_last_error reads it like any synchronous throw.

Description

~49 LOC, all in c_bridges/node-bridge.cc:

• New thread_local t_last_unhandled_rejection separate from t_last_error so a sync throw isn't clobbered.
• SetPromiseRejectCallback installed in cs_node_lazy_init right after the Environment is set up.
• Callback filters on kPromiseRejectWithNoHandler, stringifies the rejection value.
• Cleared at the start of each eval; propagated + cleared after each eval's SpinEventLoop.

Follows the same per-call-clear → run → check-after pattern as existing try_catch error handling.

Test plan

• Reads compile cleanly (visual review against existing patterns in node-bridge.cc).
• Runtime verification requires a built vendor/node/ + libnode toolchain. My worktree doesn't have it; node-bridge.cc isn't compiled in CI either (no libnode step in ci.yml). Anyone with libnode locally can run a pragma file that rejects without .catch and observe that cs_v8_last_error now returns the rejection string instead of empty.
• Follow-up test: a fixture under tests/fixtures/pragma-interpret/ that asserts a pragma file with an unhandled rejection fails with the expected message (once pragma test infra exists).

Closes #617.

Next steps / Gaps

• Dedicated fixture under a new pragma-interpret/ test category once the pragma+libnode path has a CI lane.
• kPromiseHandlerAddedAfterReject is ignored so a late .catch() doesn't trip the error — but Node's process-level unhandledRejection event is NOT forwarded; if a Node handler is attached, V8 still fires the kPromiseRejectWithNoHandler event before the handler sees it. Cleaner integration would route through process.on('unhandledRejection') instead.
• Multiple rejections in a single eval are collapsed into one — only the last is reported.

[github-actions]

Benchmark Results (Linux x86-64)

Benchmark	C	ChadScript	Go	Node	Place
Cold Start	0.9ms	0.8ms	1.2ms	27.0ms	🥇
Fibonacci	0.814s	0.763s	1.560s	3.216s	🥇
Hash Map Lookup	0.091s	0.063s	0.090s	0.113s	🥇
Binary Trees	1.375s	1.271s	2.709s	1.181s	🥈
File I/O	0.122s	0.093s	0.087s	0.208s	🥈
JSON Parse/Stringify	0.035s	0.053s	0.182s	0.135s	🥈
N-Body Simulation	1.698s	2.131s	2.202s	2.385s	🥈
Regex Match	0.016s	0.005s	0.021s	0.005s	🥈
SQLite	0.052s	0.374s	0.503s	0.417s	🥈
Monte Carlo Pi	0.389s	0.410s	0.405s	2.248s	🥉
Quicksort	0.215s	0.246s	0.213s	0.262s	🥉
Sieve of Eratosthenes	0.016s	0.028s	0.018s	0.039s	🥉
String Manipulation	0.008s	0.018s	0.017s	0.036s	🥉
Matrix Multiply	0.520s	0.729s	0.566s	0.358s	#4

CLI Tool Benchmarks

Benchmark	ChadScript	grep	node	xxd	Place
Hex Dump	0.564s	—	1.118s	0.131s	🥈
Recursive Grep	0.020s	0.010s	0.103s	—	🥈