docs(rag): close the corpus drift — findings/collision/fast-path/forge/env-ladder coverage (#596)

Seventeen-page sweep of the agent-facing KB against shipped behavior:
required covers_parent_criteria and per-AC criteria_verified reach the
QA/PM/task-tools pages (the QA docs also named non-callable pass_review/
fail_review — the MCP tools are pass/fail); collision_context lands in the
QA/gate/planning evidence docs; the possibilities matrix gets its own
architecture page + config entry; the auditor page gains its missing
waive_finding and playbook-curation verbs; git-pr-types.md is rewritten
off the long-dead is_root_pr model; PR/workspace/git-error pages stop
assuming GitHub (forge-agnostic + env-ladder semantics).

Co-authored-by: Renn F <rennf93@users.noreply.github.com>

docs/rag/roles/developer.md

@@ -48,7 +48,10 @@ open_pr(task_id)    → opens the PR, transitions to awaiting_qa
        └── QA fails → returns to needs_revision; fix + commit + open_pr again
 
 i_am_blocked(task_id, reason)  → external dependency; cell PM unblocks
-i_am_done(task_id, notes, resolved_findings?)  → batched verify + open_pr shortcut
+i_am_done(task_id, notes, resolved_findings?)  → batched verify + open_pr shortcut;
+                                  silently fast-paths straight to QA when the
+                                  possibilities matrix is armed and your work already
+                                  looks done (see "The possibilities-matrix fast path")
 unclaim(task_id)               → release a task back to the queue
 resume(task_id)                → recover after compact / restart
 i_am_idle()                    → no work in your queue right now
@@ -97,6 +100,14 @@ When toolchain matching is enabled, `i_am_done` is refused if the project's test
 
 When the architectural-conventions standard is enabled, `i_am_done` is refused on any block-level convention finding (e.g. a model defined in a router), reported with the offending `file:line` and a fix hint. A genuine false positive is cleared by committing a waiver in `.roboco/conventions.yml`.
 
+## The possibilities-matrix fast path
+
+When `ROBOCO_POSSIBILITIES_MATRIX_ENABLED` is armed, `i_am_done` checks whether your work already looks done — commits exist, the PR is open, every acceptance criterion is addressed, and no revision finding is still open. If so, it takes a fast path straight to `awaiting_qa` in one call instead of the standard multi-turn verify/journal derivation. You don't call anything different or opt in — you always just call `i_am_done(task_id, notes, resolved_findings?)`, and the fast path silently applies when it applies. The non-negotiable guards still run either way: ownership, branch pushed and not behind base, conventions, and every open finding named via `resolved_findings`. The fast path trusts the PR's own CI-green signal as the quality gate; if there's no CI signal it falls back to the local `make quality` gate, and a known-red CI refuses the fast path outright (fix CI, don't route around it) rather than shipping a broken build to QA.
+
+## Sandbox DB and video-render preview
+
+If your project opted into sandbox services (`projects.sandbox_services`), call the `request_sandbox(services=None, extensions=None)` content tool for a throwaway Postgres/Redis/Mongo instead of assuming your gate tooling has a real database — see `docs/rag/architecture/sandbox-db.md`. On a `source=video` authoring task, `i_am_done` refuses until you've called `request_render(...)` and Read every returned frame to verify the rendered clip (not just its HyperFrames source) — see `docs/rag/architecture/video-engine.md`.
+
 ## Recovering from a bounce (`needs_revision`)
 
 QA (`fail`), the in-path PR reviewer (`pr_fail`), your PM (`request_changes`), or the CEO (`ceo_reject`) can bounce your task back to `needs_revision` — and now the feedback is structured, not just a prose note. `evidence(task_id)` carries `revision_findings`: the OPEN entries from the revision-findings ledger, each with `file`/`line`/`severity`/`expected`/`actual`/`fix`. Read every one before you touch code — this is the actual code-level feedback, not a summary of it.