fix: feat: predictor re-evaluates prediction/backlog issues — evolve stale watches into targeted warnings (#588)

Add a re-evaluate-backlog step to the predictor formula between
collect-signals and analyze-and-predict. For each open prediction/backlog
issue, the predictor now reads the original context and planner comments,
extracts the assumptions that made it "watch, don't act", and re-checks
those conditions against current system state.

Three outcomes:
- CONDITIONS_CHANGED → file new prediction/unreviewed, close old as superseded
- STALE (30+ days, conditions stable) → close as prediction/actioned
- UNCHANGED_RECENT → skip (existing behavior)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

predictor/AGENTS.md

@@ -1,9 +1,9 @@
 <!-- last-reviewed: 9ec0c0221032979bd4440b9fd67f2072f1de01be -->
 # Predictor Agent
 
-**Role**: Risk oracle and opportunity spotter (the "goblin"). Runs a 3-step
-formula (preflight → collect-signals → analyze-and-predict) via interactive
-tmux Claude session (sonnet). Collects three categories of signals:
+**Role**: Risk oracle and opportunity spotter (the "goblin"). Runs a 4-step
+formula (preflight → collect-signals → re-evaluate-backlog → analyze-and-predict)
+via interactive tmux Claude session (sonnet). Collects three categories of signals:
 
 1. **Health signals** — CI pipeline trends (Woodpecker), stale issues, agent
    health (tmux sessions + logs), resource patterns (RAM, disk, load, containers)
@@ -27,9 +27,10 @@ memory check (skips if available RAM < 2000 MB).
   sources disinto project config, builds prompt with formula + Codeberg API
   reference, creates tmux session (sonnet), monitors phase file, handles crash
   recovery via `run_formula_and_monitor`
-- `formulas/run-predictor.toml` — Execution spec: three steps (preflight,
-  collect-signals, analyze-and-predict) with `needs` dependencies. Claude
-  collects signals and files prediction issues in a single interactive session
+- `formulas/run-predictor.toml` — Execution spec: four steps (preflight,
+  collect-signals, re-evaluate-backlog, analyze-and-predict) with `needs`
+  dependencies. Claude collects signals, re-evaluates watched predictions,
+  and files prediction issues in a single interactive session
 
 **Environment variables consumed**:
 - `CODEBERG_TOKEN`, `CODEBERG_REPO`, `CODEBERG_API`, `PROJECT_NAME`, `PROJECT_REPO_ROOT`
@@ -42,5 +43,6 @@ load formula + context → create tmux session → Claude collects signals
 (health: CI trends, stale issues, agent health, resources; outcomes: output
 freshness, capacity utilization, throughput; external: dependency advisories,
 ecosystem changes via web search) → dedup against existing open predictions →
+re-evaluate prediction/backlog watches (close stale, supersede changed) →
 file `prediction/unreviewed` issues → `PHASE:done`.
 The planner's Phase 1 later triages these predictions.