Recover stranded work for #359 and #379

CHANGELOG.md

@@ -2,6 +2,9 @@
 
 ## Unreleased
 
+- Added per-rig Ralph mode config via `sgt config ralph`, with a normalized Ralph state model in status/plan-state/cockpit output and explicit concurrency-accounting rules for admissible vs support-only vs duplicate lanes.
+- President and rig-activity logic now treat an enabled unmet Ralph condition as active work, emit dedicated Ralph underfill/idle intervention reasons, and keep stale `verified` / `waived` acceptance metadata from falsely terminating plan completion while Ralph remains unmet.
+- Added `test_ralph_mode_latest_main_proof.sh` plus README guidance as the explicit latest-main proof path for Ralph mode, bundling config/state, President supervision, refill recovery, and cockpit visibility coverage.
 - Mayor now auto-refreshes at runtime when the estimated AI prompt exceeds `SGT_MAYOR_AUTO_REFRESH_TOKENS` (default `150000`), using the same handoff/archive path as `sgt mayor refresh` plus a cooldown guard.
 - `sgt status` and `sgt status --json` now expose the latest Mayor prompt-budget estimate and last automatic refresh metadata.
 - `sgt context index` now writes through unique fsynced temp files before `os.replace`, avoiding concurrent rebuild corruption of `~/.sgt/context/<rig>/index.json`.

README.md

@@ -12,6 +12,7 @@ The repo includes a minimal Web UI for realtime monitoring and dispatch:
 
 - Docs / quick start: [`web/README.md`](web/README.md)
 - Default URL: `http://localhost:4747`
+- First-class deploy path: `sgt web deploy` from the repo checkout, then `sgt web status` / `sgt web verify` to confirm the live served tree really matches the repo source
 - Latest-main cockpit proof: `./test_web_cockpit_latest_main_proof.sh`
 
 ## Why SGT exists (short)
@@ -86,6 +87,7 @@ cat verification.md | sgt create blocker <rig> -
 What happens:
 - the blocker evidence is appended to the rig’s shared `SGT_CONTEXT.md`
 - a durable blocker record is created under `~/sgt/.sgt/acceptance-blockers/`
+- blocker records now carry a stable `SEVERITY_CLASS` plus `DEDUPE_KEY` so operator surfaces can separate control-plane incidents from normal still-red candidate churn and collapse repeated same-key pulses into one summary
 - the Mayor is woken immediately to re-create actionable work if the rig is otherwise idle
 - periodic mayor cycles will no longer treat an empty issue/PR board as success while the blocker remains open
 
@@ -206,6 +208,60 @@ sgt plan tick scrapegoat
 sgt wake-mayor "plan-update:scrapegoat"
 ```
 
+## Ralph Mode
+
+Ralph mode pins a rig into ongoing work until an explicit operator-owned Ralph condition is satisfied or the operator disables Ralph mode.
+
+Configure it per rig with:
+
+```bash
+sgt config ralph pmkb \
+  --enable \
+  --condition "1K PNL for 5m btc pipeline" \
+  --target 5
+```
+
+Optional knobs:
+
+```bash
+sgt config ralph pmkb --condition-status met
+sgt config ralph pmkb --count-support yes
+sgt config ralph pmkb --disable
+```
+
+Current model:
+- Ralph config lives in `~/.sgt/rig-config/<rig>.json` under `ralph_mode`.
+- Ralph runtime state is computed from rig config plus live rig work and is also persisted into `~/.sgt/plan-state/<rig>.json` under `ralph`.
+- While Ralph is enabled and `condition_status` is not `met` or `waived`, plan completion is forced back to a non-terminal pending rollup (`ralph-condition-unmet` or `ralph-underfilled`) even if stale acceptance metadata says `verified` or `waived`.
+- `sgt status`, `sgt status --json`, and the Web cockpit expose Ralph state, condition, target concurrency, active lane count, admissible lane count, underfill state, and the latest President action for that rig.
+- President treats Ralph underfill with existing admissible backlog as a bounded refill wake, but upgrades idle, fake-complete, or no-refillable-backlog Ralph states into a stronger refresh/contradiction intervention so the rig does not silently drift.
+
+Concurrency accounting policy:
+- A Ralph lane is a distinct open `sgt-authorized` issue number on that rig.
+- Issues labeled `duplicate` or `redundant` never count toward Ralph concurrency.
+- Issues labeled `support-only` or `cleanup-only` do not count by default; opt in with `sgt config ralph <rig> --count-support yes`.
+- `active_lane_count` counts admissible lanes that currently have a live polecat on that issue.
+- `backlog_lane_count` counts admissible open lanes that exist but are not currently active.
+
+Example:
+- Issue `#101` with a live polecat counts as one active Ralph lane.
+- Issue `#102` labeled `support-only` does not count unless support lanes are enabled.
+- Issue `#103` labeled `duplicate` never counts.
+
+### Ralph latest-main proof path
+
+Run this on a fresh checkout of the latest `master`/mainline commit when you want repo-owned proof that Ralph mode configuration, state exposure, concurrency accounting, and bounded refill behavior still work end to end:
+
+```bash
+./test_ralph_mode_latest_main_proof.sh
+```
+
+That proof path bundles the checks that matter for Ralph mode:
+- `test_ralph_mode_config_and_state.sh` proves per-rig Ralph config, normalized state output, anti-false-completion semantics, and admissible/support/duplicate lane accounting.
+- `test_president_runtime_supervision.sh` proves President distinguishes Ralph underfill, idle, and contradiction states and picks the expected bounded intervention.
+- `test_mayor_stranded_zero_worker_recovery.sh` proves controller-side refill logic can top an actionable rig back up toward target live concurrency instead of leaving it quietly idle.
+- `web/test/cockpit.test.js` proves the Web cockpit snapshot keeps Ralph state attached to each rig for operator visibility.
+
 ## CI self-healing (Mayor watchdog)
 
 For rigs with `SGT_PLAN.json`, Mayor also watches the latest **master** GitHub Actions run.