fix: keep the dev shell out of the daily driver's runtime file

[omartelo]

The bug

runChromium records runtime.json before the LICH_DEV_URL branch, so every task dev run wrote its own port and
token into the daily driver's runtime file — and deleted that file on its clean exit (defer os.Remove(path)). Two
casualties: install.sh could no longer reach the running daily driver for /restart, and a second daily-driver
launch stopped recognising itself as a duplicate, dying with a bare bind error instead of focusing the open window.

The fix

singleton.path now gates the filename on LICH_DEV, mirroring store.databasePath and logging.fileName:
runtime-dev.json for a task dev shell, runtime.json for the daily driver.

The gate lives at the shared path rather than at the call site because Write, Read and Detect all route through
path — one guard covers the write, the read and the exit-time removal, and the dev shell's duplicate detection now
finds its own instance (on 47822) instead of reading the daily driver's record and mistaking a live app for a dead
port. install.sh keeps reading runtime.json and keeps finding the real instance.

No CHANGELOG.md entry: the bug is reachable only through task dev, so no user of a published version lived it.

Test plan

• TestDevSplit covers both halves of the bug: a dev Write leaves the daily record intact, and removing the path
  dev was handed (what main.go defers) leaves runtime.json standing.
• go test ./... — 405 passed in 19 packages.
• go vet ./... and gofmt -l . clean.
• Cross-compile loop for linux, darwin and windows builds and vets clean.