Add bidirectional RF handoff coordination

docs/handoff/TX_AGENT_USAGE.md

@@ -0,0 +1,47 @@
+# MAIN TX watcher usage
+
+## Purpose
+
+tools/tx_agent_watcher.py is an ordinary local Python process. It polls the small RX result on origin/rx-node. It does not call an LLM and does not open HackRF hardware.
+
+## Start
+
+From the repository root, double-click:
+
+    tools\start_tx_watcher.cmd
+
+The visible console polls every 5 seconds. A missing rx-node branch or invalid result is reported. No model runs while polling.
+
+## READY and RED approval
+
+For a fresh valid RF001A_OTA_CAPTURE_002 READY result, the watcher prints the task ID, capture start and expected end. It then asks for:
+
+    AUTHORIZE RF TX
+
+Any other input blocks authorization.
+
+After the exact phrase, the watcher fetches rx-node again. It requires the same result digest, a fresh publication, an active 120-second capture window and at least 10 seconds remaining.
+
+On success it writes:
+
+    .local\tx_authorizations\RF001A_OTA_TX_002_authorized.json
+
+This record opens the gate for the separately controlled TX executor. The watcher itself does not run hackrf_transfer. The record explicitly says rf_tx_started = false and rf_tx = 0.
+
+The TX executor must re-fetch rx-node and compare the result commit and digest immediately before hardware access.
+
+## Stop
+
+Double-click:
+
+    tools\stop_tx_watcher.cmd
+
+The stop command writes a local stop request and waits for the watcher PID to clear.
+
+## Fake mode
+
+Automated tests use a local bare Git repository. For a single local JSON fixture, the CLI dry-run form is:
+
+    python -B tools/tx_agent_watcher.py --dry-run --once --result-file <ready-json> --approval authorize
+
+The --approval option is rejected outside --dry-run. Fake mode writes a local fake authorization and never fetches the production remote or accesses SDR hardware.