feat(operator): introduce v0.1 local web UI - #39

Translation:

2. Shutdown can miss a run that is still starting.
   A run starts in the background. If you shut the operator down at exactly that moment, shutdown may not know the background
   startup exists yet. That background work then continues after shutdown and tries to use a closed operator.
   Fix:make startup and shutdown coordinate so shutdown always waits for—or cancels—the startup work.
3. A node can show logs from the wrong attempt.
   If the same agent/node runs twice, opening the details for its second run can include steps and evidence from its first run.
   Fix:show only data belonging to the specific run attempt the user selected.
4. Restarting the operator can leave the UI showing old workflows.
   The UI uses a revision number to decide whether workflow data is newer. After an operator restart, revisions restart from a
   small number, so the UI can reject the fresh data and keep data from the previous operator process.
   Fix:when the operator itself changes, discard the old catalog regardless of revision number.
5. The web interface handles bad input and huge histories badly.

• Sending corrupt protocol data returns “server crashed” instead of “your request is invalid.”
• If a live stream fails halfway through, the server tries to send a second response after it already started the first.
• When loading run history, the browser keeps fetching every page indefinitely. A large run history could consume all browser
  memory.
  Fix:return the correct client-error response, handle stream errors in-place, and cap how much historical data the browser
  loads.

6. The migration is incomplete and CI does not protect the new frontend.
   The PR introduces a new workflow-selection API but leaves the obsolete API path in place. Documentation still names old APIs.
   CI also does not run the web build/tests, so broken frontend code or stale packaged web assets can merge unnoticed.
   Fix:delete the old API path, update docs, and make CI build/test/check generated frontend assets.

Requested changes addressed

The review was performed at 6f8d489; the current head is 43bf408. Here is how each finding was resolved.

1. Branch test failures / lifecycle contract

The new asynchronous lifecycle is now explicit: start_run() publishes a REQUESTING run immediately, while preparation/build/protocol failures terminalize that retained run as FAILED with an operator log rather than escaping synchronously from start_run(). The affected preparation, reload, state-detail, and packaged-asset assertions were updated to test that observable contract. The current head is green across Lint + tests, Browser UI, Ray tests, and Tmux TUI tests.

2. close() racing asynchronous preparation

Preparation-thread ownership is installed while holding the operator lock, before the thread can run. Startup rechecks the closed state at the synchronization boundary, and shutdown cancels/drains active work and joins the owned preparation thread. A close that overlaps REQUESTING can therefore no longer miss a not-yet-published thread or allow background preparation to publish into a closed operator.

3. Trace contamination across agent invocations

Finalized trace reconstruction is now scoped by invocation_id. Steps and evidence for a retained node trace are selected only from the invocation represented by that trace header, so a later invocation cannot inherit iterations or evidence from an earlier invocation of the same node. The regression coverage uses distinct invocations and verifies the reconstructed content/counts stay aligned.

4. Catalog revisions across operator restarts

Catalog monotonicity is now evaluated within an operator_instance_id epoch. A new operator instance always replaces the previous catalog regardless of its lower revision, while stale revisions are still rejected within the same epoch. Reset/baseline installation and catalog callbacks now carry that epoch consistently, preventing old-process workflows from surviving a restart.

5. gRPC-Web errors and bounded browser hydration

• Malformed protobuf payloads, including DecodeError, are translated to INVALID_ARGUMENT instead of an internal server error.
• Streaming failures are translated inside the already-started chunked response and emitted as gRPC-Web trailers; the handler no longer attempts a second HTTP response. Response iterators are closed during stream teardown/disconnect handling.
• Browser baseline loading is bounded before publication: at most 100 pages, 10,000 run summaries, and 8 MiB of encoded summaries. Repeated/non-advancing page tokens fail immediately, and no partial baseline is installed.
• gRPC-Web routing is derived from the generated service descriptor rather than a hand-maintained RPC list.

6. V2 cutover and frontend CI protection

• The legacy flow_name request field, dual-population fallback, and client-side legacy-name cache were removed. workflow_selector is now the single run/list selector, with protobuf bindings, callers, tests, and architecture documentation updated together.
• CI now regenerates the TypeScript client, runs all browser tests, builds the packaged assets, executes the volume and real-Chrome benchmark, and checks generated/protobuf/asset trees for a clean diff.
• The real-browser job uses browser-actions/setup-chrome@v2. Its remaining failure was traced to the benchmark fixture importing the removed /src/tailwind.css; it now imports /src/style.css. The same Chrome-for-Testing 151 binary passed locally with 10,000 runs, 14 DOM rows, 102.90 ms render, and 41.50 ms interaction, and the current Browser UI Actions job is green.

Owner-requested cleanup

The interim CI fixes were rewritten into one commit, 0d16698 (ci: run browser benchmark with Chrome v2). The OpenSpec planning directory was removed in 43bf408 and is no longer present in the PR tree.

Current verification

• Lint + tests
• Browser UI
• Ray tests
• Tmux TUI tests