feat(dev): bring run_inner to parity with dev-runner.sh — log shipper on the live path

[I-am-nothing]

The compose runner ran scripts/dev-runner.sh (air hot-reload, dev-proxy, livereload, per-module token files — but no log shipper), while the Rust run_inner had the #36 shipper but none of the dev features, so it was never the live path and the developer Logcat stayed empty.

run_inner now carries the full runner:

• Polling hot-reload (new reload.rs): mtime scan of .go/.sql every 2s (excludes tmp/web/node_modules — bind mounts deliver no inotify), go build -buildvcs=false -o <tmp>/<slug> ., restart on change, keep the old process on a failed rebuild, and re-tap the new child's stdout/stderr into the forwarder + shipper on every restart.
• In-process /_m/<slug>/* dev-proxy (new proxy.rs) on PROXY_PORT (default 8080): TrimPrefix semantics (empty→/, query preserved), raw duplex relay after the rewritten head so chunked/SSE/upgrades pass through; backend connect refused → 502 (startup race).
• Deterministic PORT from 18080 in go.work order; livereload LR_PORT from 8089 per web-enabled module.
• Per-module MS_PLATFORM_TOKEN_FILE=<root>/.ms-platform-token-<slug> (distinct token per tunnel; a shared file would 401 all but the first).
• esbuild web watcher: npm install --silent then node esbuild.config.mjs --watch, output prefixed [<slug>:web].
• --watch is honored (was dead surface: run_inner(root, _args)); defaults on so the compose command stays mirrorstack dev --all, --watch=false gives the old one-shot behavior.
• One-line warning when MS_DISPATCH_URL/MS_INTERNAL_SECRET are unset so a config-disabled Logcat is not silent.
• Shutdown tears down supervisors, module children, and esbuild watchers, then lets the shipper threads flush their final batch.

log_shipper.rs is untouched — putting run_inner on the live path is the point of mirrorstack-cli#38.

Validation

cargo check, cargo clippy --all-targets --all-features, and cargo fmt --check clean; 151 tests pass.

Closes #38

🤖 Generated with Claude Code