e2e: standard bridge conversation test harness

A repeatable multi-turn primary↔worker conversation driven entirely through the
bridge's loopback REST face (async fire-and-poll) — never sets ANTHROPIC_BASE_URL
and never touches herdr, so it is subscription-safe by construction. Records every
turn to a transcript, grades each (OK / DEGRADED / EMPTY / FAILED / WEDGE), and
exits non-zero if any turn fails to deliver-and-reply, so it is CI-usable.

This is the repeatable form of the ad-hoc channel test that surfaced the CB-115 and
CB-116 gaps.

e2e/README.md

@@ -0,0 +1,77 @@
+# Bridge conversation test (`e2e/`)
+
+A standard, repeatable **live** end-to-end test of the two-way channel: it drives a real
+multi-turn conversation between a primary and an off-subscription worker **through the
+running `bridged` daemon**, captures the full transcript, and grades the channel.
+
+This is the committed form of the ad-hoc channel test that discovered the CB-115 gaps
+(herdr `unknown` misclassification wedging delivery, dirty completion scrapes, and workers
+never calling `bridge_reply` in conversation). Run it after any change to the injector,
+status handling, completion/failure paths, or the worker reply charter.
+
+## What it exercises
+
+Each turn goes through the whole gateway exactly as a primary Opus session would — async
+fire-and-poll (`POST /sessions/{id}/message {"wait":false}` → `GET /tasks/{ticket}`), so it
+also validates the path that beats the caller's MCP timeout. It never sets
+`ANTHROPIC_BASE_URL` and never talks to herdr directly, so it is **subscription-safe by
+construction** — it only calls the bridge's loopback REST face.
+
+```mermaid
+sequenceDiagram
+    participant T as conversation_test.py
+    participant B as bridged (REST)
+    participant W as worker (off-sub)
+    T->>B: POST /workers  (spawn)
+    T->>B: GET /sessions/{id}/status  (await ready)
+    loop each turn
+        T->>B: POST /sessions/{id}/message {wait:false}
+        B-->>T: ticket
+        B->>W: inject prompt (status-gated)
+        W-->>B: bridge_reply
+        T->>B: GET /tasks/{ticket}  (poll)
+        B-->>T: done + reply
+    end
+    T->>B: DELETE /workers/{pane}  (stop)
+```
+
+## Prerequisites
+
+- `bridged` is running (default REST on `http://127.0.0.1:8765`) with at least one worker
+  profile configured and its backend reachable.
+- herdr is up (the daemon needs it).
+- Python 3 (standard library only — no pip installs).
+
+## Run
+
+```bash
+# spawn the default-profile worker, run the built-in 5-turn conversation, grade, clean up
+python3 e2e/conversation_test.py
+
+# pick a profile / reuse a live worker / use your own prompts
+python3 e2e/conversation_test.py --profile ollama
+python3 e2e/conversation_test.py --tid term_abc123 --keep-worker
+python3 e2e/conversation_test.py --prompts my_prompts.txt --out /tmp/run1
+```
+
+A prompts file is one prompt per line; blank lines and `#` comments are ignored.
+
+## Output & grading
+
+- Writes `transcript.md` (in `--out`, default `e2e/`) — every turn's prompt, worker reply,
+  latency, resolution source, and observed status transitions, with an inline `> **GAP**`
+  note on any non-clean turn.
+- Prints a per-turn line and an overall summary, and **exits non-zero** if any turn wedged,
+  failed, or returned empty — so it is CI-usable.
+
+Per-turn grade:
+
+| Grade      | Meaning                                                              |
+|------------|---------------------------------------------------------------------|
+| `OK`       | delivered and resolved by an explicit `bridge_reply` (`source=reply`) |
+| `DEGRADED` | delivered and answered, but resolved via completion-scrape fallback |
+| `EMPTY`    | turn completed but the reply was empty                              |
+| `FAILED`   | the worker's turn ended in failure (`phase=failed`)                 |
+| `WEDGE`    | never resolved within the poll window (delivery wedge / lost turn)  |
+
+`PASS` requires every turn to be `OK` or `DEGRADED`; a clean run is every turn `OK`.