flue dev crashes with EMFILE when system hits file watcher limits

[astrobot-houston]

Originally reported by @kalepail in #313.

All community-submitted pull requests are automatically converted to issues (bugs) & discussions (feature requests, enhancements) where they can be triaged and prioritized. Once prioritized, a PR implementation is created automatically.

Describe the Bug

Running flue dev in a large repository or on a system with file watcher pressure causes the process to crash before the local dev server starts. The crash occurs because fs.watch(projectRoot) throws EMFILE: too many open files, watch (or ENOSPC) when the system's file watcher limit is exhausted, and the error is unhandled.

This is unnecessary because Flue already has the exact list of resolved config files it needs to watch and could poll them individually instead of watching the entire parent directory.

Expected Behavior

flue dev should start the local server even when the system cannot watch the config directory due to resource exhaustion (EMFILE/ENOSPC). Config file change detection should fall back to polling the resolved config files individually, and non-resource watcher errors should continue to fail fast.

Steps to Reproduce

1. Open a large repository (or a system under file watcher pressure where the OS watcher limit is nearly exhausted)
2. Run flue dev
3. Observe that the process exits with EMFILE: too many open files, watch before the dev server starts

---

Original implementation from #313 by @kalepail

[kalepail]

Adding the concrete implementation/test details from the closed PR so they remain visible on the triage issue.

The proposed patch in #313 changed packages/cli/bin/flue.ts to:

• catch synchronous fs.watch(directory) failures for config directories;
• catch async watcher error events;
• treat only EMFILE / ENOSPC as recoverable watcher-resource exhaustion;
• fall back to fs.watchFile() polling of the already-resolved config file list;
• preserve existing fail-fast behavior for unrelated watcher errors.

The regression test added to packages/cli/test/dev.test.mjs:

• preloads a small shim that makes fs.watch(process.cwd()) throw EMFILE;
• runs flue dev;
• verifies the server still starts;
• edits flue.config.mjs;
• verifies the dev session restarts and emits the new build output.

Validation run on the PR branch:

pnpm --filter @flue/runtime --filter @flue/sdk --filter @flue/cli run build
pnpm --dir packages/cli run check:types
pnpm --dir packages/cli exec node --test test/dev.test.mjs

All passed locally.

One nuance: this issue is specifically the parent config watcher. There is a related but broader dev-watch concern around generated directories / large result directories causing rebuild churn. That is probably better handled as a separate configurable watch-ignore / .gitignore-aware enhancement, not bundled into the config-watch fallback fix.