Merge pull request 'fix(§22/registry): prune projects absent from the registry on reconcile (v0.55.0)' (#51) from fix-provision-reconcile into main

The registry mirror was additive-only — it upserted the projects/collections a
deployment's projects.yaml declares but never removed ones it had dropped.
Re-pinning to a different registry stranded the old project's collections + its
cached entries as dead rows that, at scale, starved writes (PPE accumulated
~1.2k orphaned entry rows and had to be reset by hand).

`registry.refresh_registry(prune=True)` now calls `projects.prune_absent_projects`,
deleting a removed project with its collections and every project-scoped row
(the 7 project_id-keyed tables + 13 collection_id-keyed entry-corpus tables +
the non-keyed thread_messages descendant) in one FK-off transaction with a
`PRAGMA foreign_key_check` backstop that rolls back rather than leave a dangling
ref — mirroring restamp_default_project / reconcile_default_collection_id.

Scoped for safety: whole-project granularity (a present project is untouched);
the default project is never pruned; prune runs ONLY on the full-reconcile
triggers (startup + the periodic sweep), never on incidental refreshes (collection
create, webhook), so an in-app refresh can't wipe a project. Reached only after a
successful parse of a non-empty registry (parse_registry rejects empty; the read
raises on transport error), so a transient read can't trigger a wipe.

Second of the two §9-surfaced framework fragilities. No migration. backend 685
green (+5: prune removes-all/no-op/never-default/empty-rejected/incidental-no-prune
with FK-integrity backstop).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -23,6 +23,95 @@ skip versions are the composition of each intervening adjacent
 release's steps in order — no A-to-B path is pre-computed beyond
 that.
 
+## 0.55.0 — 2026-06-09
+
+**Minor — the registry reconcile now prunes projects the registry no longer
+declares; read the upgrade note.**
+
+The registry mirror was additive-only: it upserted the projects/collections a
+deployment's `projects.yaml` declares but never removed ones it had dropped.
+Re-pinning a deployment to a different registry (or editing `projects.yaml` to
+remove a project) therefore stranded the old project's collections and its
+cached entries in the database — dead rows that, at scale, starved writes (a
+deployment's PPE accumulated ~1,200 orphaned entry rows this way and had to be
+reset by hand).
+
+- **Prune-on-reconcile.** `registry.refresh_registry(..., prune=True)` now calls
+  `projects.prune_absent_projects`, which deletes any project absent from the
+  freshly-parsed registry together with its collections and every project-scoped
+  row (entries + their satellites, threads + messages, changes, PRs, stars,
+  watches, …) in one transaction, with FK enforcement off for the multi-table
+  delete and a `PRAGMA foreign_key_check` backstop that rolls the whole thing
+  back rather than leave a dangling reference.
+- **Whole-project granularity, and only the default-safe triggers.** A project
+  still present in the registry is never touched here (pruning individual entries
+  within a present project belongs to the best-effort content cache). The
+  deployment's **default project is never pruned**, whatever the registry says.
+  Pruning runs only on the two full-reconcile triggers — **startup** and the
+  **periodic reconciler sweep** — never on incidental refreshes (a collection
+  create, a registry webhook), so an in-app refresh can't wipe a project.
+- **Safe against a transient registry read.** The prune is reached only after a
+  successful parse of a non-empty registry: `parse_registry` rejects an empty
+  project list and the registry read raises on a missing file / transport error
+  (the reconciler keeps last-good on that error). A flaky read can't trigger a
+  wipe.
+
+**Upgrade note (no steps required for a correctly-configured deployment).** On
+the first startup after upgrading, any project that exists in the database but is
+**absent from your `projects.yaml`** will be pruned. For a deployment whose
+database was only ever populated from its current registry this is a no-op (every
+DB project is in `projects.yaml`). If you have deliberately out-of-band projects
+in the database (there is no supported way to create one — every project goes
+through `projects.yaml`), add them to the registry before upgrading. No database
+migration and no configuration change.
+
+## 0.54.1 — 2026-06-09
+
+**Patch — the Gitea-OAuth sign-in no longer 500s when an account was first
+created via OTC; no operator action required.**
+
+`auth.provision_user` (the Gitea-OAuth callback path) matched an existing
+`users` row only by `gitea_id`. A human who signed in first via the OTC
+email path owns an email-only row (`gitea_id` NULL); when they later signed
+in via Gitea OAuth with the same email, the lookup missed and the code
+INSERTed a new row — colliding on the `idx_users_email` case-insensitive
+unique index and raising `IntegrityError`, which 500-ed the callback. Now,
+when no `gitea_id` row exists, `provision_user` reconciles by email and links
+the OAuth identity (`gitea_id` / `gitea_login` / profile) onto the existing
+row — the mirror of the OTC linker, which already links an OAuth-era email row
+on first OTC sign-in. The §6.1 owner-zero bootstrap is applied on link
+(matching the fresh-insert path); the row's `permission_state` is preserved so
+linking never changes a human's admission status as a side effect.
+
+## 0.54.0 — 2026-06-09
+
+**Minor — the AI "Ask" affordance now works from an entry's canonical view
+(§8.12); no operator action required.**
+
+AI chat is an editing activity (a turn can emit proposed edits) and only runs
+on an edit branch. On an entry's canonical `main` view the right column is the
+human-discussion panel, not the chat panel — so invoking "Ask" (the selection
+tooltip or the prompt bar) from the reading view pushed a chat turn that had
+nowhere to render: it silently did nothing. Surfaced by dogfooding a non-default
+project's spec entry on a live deployment.
+
+- **Asking while reading "just works" (Option B).** When Ask is invoked from
+  the canonical view, the app transparently cuts an edit branch via the same
+  dispatch as Start Contributing (`start-edit-branch` for a super-draft,
+  `promote-to-branch` for an active RFC — idempotent, so it reuses an existing
+  edit branch rather than double-cutting), navigates onto that branch so the
+  chat panel mounts, and runs the question (text + the selected quote) there
+  once the branch's chat thread has resolved. The viewer lands in the chat with
+  the streaming answer, quote intact.
+- **Graceful degradation.** A signed-out viewer keeps the read-only path
+  (sign-in prompt; Ask disabled) and a viewer who cannot contribute has the
+  branch cut rejected server-side and the error surfaced — no silent no-op and
+  no spurious branch. Asking from a branch already, and Flag on the canonical
+  view (which still opens a human discussion thread), are unchanged.
+
+Frontend-only: the branch chat endpoints already worked on a branch. No
+database migration and no configuration change.
+
 ## 0.53.0 — 2026-06-09
 
 **Minor — the branch/edit/body subsystem is now three-tier (project /

SPEC.md

@@ -1022,6 +1022,20 @@ the user on it in contribute mode. New-branch naming defaults to an
 auto-generated value (user-renamable); the exact format is an
 implementation detail.
 
+The AI chat is an editing activity — a turn can emit `<change>`
+proposals — so it runs on an edit branch, not on read-only main, whose
+right column is the human-discussion surface (§8.12 is branch-scoped).
+Invoking the AI **Ask** affordance (the selection tooltip's prompt, or
+the prompt bar) from main therefore transparently cuts an edit branch
+via the same dispatch as "Start Contributing" (promote-to-branch for an
+active RFC, start-edit-branch for a super-draft per §9.5; idempotent, so
+an existing edit branch is reused rather than a second one cut), lands
+the user on it, and runs the question — text plus any selected quote —
+as the branch's first chat turn. A viewer who cannot contribute has the
+branch cut rejected and the error surfaced (no branch is created); a
+signed-out viewer keeps the §8.7 read-only path. **Flag** from main is
+unaffected — it opens a human discussion thread on main, not a branch.
+
 Discuss vs. contribute is an *intent* affordance, not a *permission*
 affordance. A user without contribute access to a branch sees the
 toggle disabled, with a sign-in or request-access path (see §8.7).

VERSION

@@ -1 +1 @@
-0.53.0
+0.55.0

backend/app/auth.py

@@ -145,6 +145,19 @@ def provision_user(config: Config, profile: dict[str, Any]) -> SessionUser:
 
     c = db.conn()
     existing = c.execute("SELECT * FROM users WHERE gitea_id = ?", (gitea_id,)).fetchone()
+    # No row for this gitea_id yet. A prior OTC sign-in (`provision_or_link_user`)
+    # may have created an email-only row with this same email and `gitea_id`
+    # NULL; `idx_users_email` is a unique index, so a blind INSERT below would
+    # raise IntegrityError and 500 the OAuth callback. Reconcile by email: link
+    # the OAuth identity onto that existing row instead. Mirror of the OTC
+    # linker, which links an OAuth-era email row on first OTC sign-in.
+    linked = False
+    if existing is None and email:
+        existing = c.execute(
+            "SELECT * FROM users WHERE email = ? COLLATE NOCASE AND gitea_id IS NULL LIMIT 1",
+            (email,),
+        ).fetchone()
+        linked = existing is not None
     if existing is None:
         role = "owner" if config.owner_gitea_login and login == config.owner_gitea_login else "contributor"
         # v0.8.0: a fresh OAuth-provisioned user is also subject to
@@ -166,15 +179,26 @@ def provision_user(config: Config, profile: dict[str, Any]) -> SessionUser:
         permission_state = "granted"
     else:
         user_id = existing["id"]
-        role = existing["role"]
+        # On a fresh link this OAuth sign-in is the row's first, so apply the
+        # §6.1 owner-zero bootstrap (matching the INSERT path). For a row already
+        # matched by gitea_id the role is settled — preserve it. `permission_state`
+        # is preserved either way: linking an OAuth identity onto an existing
+        # email row must not silently change the human's admission status.
+        if linked and config.owner_gitea_login and login == config.owner_gitea_login:
+            role = "owner"
+        else:
+            role = existing["role"]
         permission_state = existing["permission_state"] or "granted"
+        # Setting `gitea_id` matters only on the link path (it was NULL); on the
+        # gitea_id-matched path it re-writes the same value. `role` is likewise a
+        # no-op there. Always refresh the mutable profile fields + last_seen.
         c.execute(
             """
             UPDATE users
-               SET gitea_login = ?, email = ?, display_name = ?, avatar_url = ?, last_seen_at = datetime('now')
+               SET gitea_id = ?, gitea_login = ?, email = ?, display_name = ?, avatar_url = ?, role = ?, last_seen_at = datetime('now')
              WHERE id = ?
             """,
-            (login, email, display, avatar, user_id),
+            (gitea_id, login, email, display, avatar, role, user_id),
         )
 
     return SessionUser(

backend/app/cache.py

@@ -798,7 +798,7 @@ class Reconciler:
         log.info("reconciler: starting sweep")
         try:
             try:
-                await registry_mod.refresh_registry(self._config, self._gitea)
+                await registry_mod.refresh_registry(self._config, self._gitea, prune=True)
             except Exception:
                 log.exception("reconciler: registry refresh failed; keeping last-good projects")
             await refresh_meta_repo(self._config, self._gitea)

backend/app/main.py

@@ -133,7 +133,7 @@ async def lifespan(app: FastAPI):
     # (loud-fail per separation-of-concerns); the reconciler sweep keeps it
     # fresh thereafter and tolerates a later bad PR.
     try:
-        await registry_mod.refresh_registry(config, gitea)
+        await registry_mod.refresh_registry(config, gitea, prune=True)
     except Exception as e:
         raise RuntimeError(
             f"registry mirror failed at startup ({config.registry_repo_full}/projects.yaml): {e}"

backend/app/projects.py

@@ -181,6 +181,98 @@ def reconcile_default_collection_id(config: Config) -> None:
     )
 
 
+def prune_absent_projects(config: Config, present_ids: set[str]) -> list[str]:
+    """Delete projects the registry no longer declares, plus their collections
+    and every project-scoped row. Whole-project granularity only.
+
+    `present_ids` is the authoritative set of project ids from the just-parsed
+    registry. Safety is structural: the only caller (`registry.refresh_registry`)
+    raises on a read/transport error, and `parse_registry` rejects an empty
+    project list — so this is reached ONLY after a real, non-empty registry has
+    been parsed. A transient registry-read failure therefore can't trigger a wipe.
+    The default project is never pruned regardless of what the registry says.
+
+    Whole-project, not per-collection/entry: a project still present keeps all its
+    rows here. Pruning individual entries within a present project belongs to the
+    best-effort content cache (whose list can transiently fail), not to this.
+
+    Mirrors `restamp_default_project`'s FK-off + `foreign_key_check` pattern. After
+    029 the entry-corpus tables key on `collection_id` (not `project_id`), so the
+    delete spans three discovered sets — project_id-keyed tables, collection_id-
+    keyed tables (keyed to the stale projects' collections), and the lone non-keyed
+    descendant `thread_messages` (FK→threads) — in one transaction with FK
+    enforcement off and a `foreign_key_check` backstop that rolls back on any
+    dangling reference (so an incomplete delete fails loudly rather than corrupts).
+    Returns the pruned project ids (empty when nothing is stale).
+    """
+    keep = set(present_ids)
+    keep.add(resolved_default_id(config))
+    keep.add(DEFAULT_PROJECT_ID)
+    conn = db.conn()
+    stale = [
+        r["id"] for r in conn.execute("SELECT id FROM projects").fetchall()
+        if r["id"] not in keep
+    ]
+    if not stale:
+        return []
+
+    stale_collections = [
+        r["id"] for r in conn.execute(
+            f"SELECT id FROM collections WHERE project_id IN ({','.join('?' * len(stale))})",
+            stale,
+        ).fetchall()
+    ]
+    tables = [r["name"] for r in conn.execute("SELECT name FROM sqlite_master WHERE type='table'")]
+
+    def _has(table: str, col: str) -> bool:
+        return any(c["name"] == col for c in conn.execute(f"PRAGMA table_info({table})"))
+
+    pid_tables = [t for t in tables if t != "projects" and _has(t, "project_id")]
+    cid_tables = [t for t in tables if _has(t, "collection_id")]
+    p_ph = ",".join("?" * len(stale))
+    c_ph = ",".join("?" * len(stale_collections)) if stale_collections else ""
+
+    conn.execute("PRAGMA foreign_keys = OFF")
+    try:
+        conn.execute("BEGIN")
+        # Non-keyed descendant: thread_messages hangs off threads(id), which is
+        # project_id-keyed below. Clear it first by thread lineage.
+        conn.execute(
+            f"DELETE FROM thread_messages WHERE thread_id IN "
+            f"(SELECT id FROM threads WHERE project_id IN ({p_ph}))",
+            stale,
+        )
+        if stale_collections:
+            for t in cid_tables:
+                conn.execute(
+                    f"DELETE FROM {t} WHERE collection_id IN ({c_ph})", stale_collections
+                )
+        for t in pid_tables:  # includes collections + cached_prs + threads/changes/…
+            conn.execute(f"DELETE FROM {t} WHERE project_id IN ({p_ph})", stale)
+        conn.execute(f"DELETE FROM projects WHERE id IN ({p_ph})", stale)
+        violations = conn.execute("PRAGMA foreign_key_check").fetchall()
+        if violations:
+            conn.execute("ROLLBACK")
+            raise RuntimeError(
+                f"prune left foreign-key violations: {[tuple(v) for v in violations]}"
+            )
+        conn.execute("COMMIT")
+    except Exception:
+        try:
+            conn.execute("ROLLBACK")
+        except Exception:
+            pass
+        raise
+    finally:
+        conn.execute("PRAGMA foreign_keys = ON")
+    log.warning(
+        "prune: removed %d project(s) absent from the registry: %s "
+        "(%d collection(s), across %d project- + %d collection-keyed tables)",
+        len(stale), stale, len(stale_collections), len(pid_tables), len(cid_tables),
+    )
+    return stale
+
+
 def default_content_repo(config: Config) -> str | None:
     """The content repo the single-corpus mirror reads, from the default
     project's row (filled by the registry mirror). Replaces the retired

backend/app/registry.py

@@ -333,11 +333,18 @@ async def _mirror_named_collections(config: Config, gitea: Gitea, doc: RegistryD
             _upsert_named_collection(proj, subdir, ce, sha)
 
 
-async def refresh_registry(config: Config, gitea: Gitea) -> None:
+async def refresh_registry(config: Config, gitea: Gitea, *, prune: bool = False) -> None:
     """Mirror REGISTRY_REPO/projects.yaml into projects + deployment.
 
     Idempotent. Raises RegistryError on a missing/invalid file and GiteaError
     on transport failure; the caller chooses fatal-vs-tolerated.
+
+    `prune=True` additionally removes projects the registry no longer declares
+    (and their collections + entries) — see `projects.prune_absent_projects`.
+    It is OFF by default and enabled only on the full-reconcile triggers
+    (startup + the periodic sweep), never on incidental refreshes (a collection
+    create, a webhook) that don't change the project set — keeping the
+    destructive sweep on the few code paths that mean "reconcile to the registry."
     """
     item = await gitea.get_contents(
         config.gitea_org, config.registry_repo, "projects.yaml", ref="main"
@@ -355,4 +362,18 @@ async def refresh_registry(config: Config, gitea: Gitea) -> None:
     apply_registry(doc, sha, projects_mod.resolved_default_id(config))
     # §22 S2: discover + upsert named collections from each content repo.
     await _mirror_named_collections(config, gitea, doc, sha)
-    log.info("registry: mirrored %d project(s) at %s", len(doc.projects), sha)
+    # Prune projects the registry no longer declares (additive-only was a §9
+    # fragility — a re-pin to a new registry stranded the old projects' entries,
+    # starving writes). Safe here: we only reach this line after a successful
+    # parse of a non-empty registry (parse_registry rejects an empty project list
+    # and the read above raises on transport failure), so a transient failure
+    # never prunes. Whole-project granularity; the default project is never pruned.
+    pruned = (
+        projects_mod.prune_absent_projects(config, {e.id for e in doc.projects})
+        if prune else []
+    )
+    log.info(
+        "registry: mirrored %d project(s) at %s%s",
+        len(doc.projects), sha,
+        f"; pruned {len(pruned)} absent: {pruned}" if pruned else "",
+    )

backend/tests/test_provision_user_email_link.py

@@ -0,0 +1,107 @@
+"""§6 hardening — the Gitea-OAuth `auth.provision_user` reconciles by email.
+
+Regression cover for a §9-surfaced fragility: a human who signed in first via
+OTC owns an email-only `users` row (`gitea_id` NULL). When they later sign in
+via Gitea OAuth with the SAME email, `provision_user` used to match only by
+`gitea_id`, miss the OTC row, and INSERT a new row — colliding on the
+`idx_users_email` unique index and 500-ing the callback. The fix links the OAuth
+identity onto the existing email row (the mirror of the OTC linker).
+"""
+from __future__ import annotations
+
+from test_propose_vertical import (  # noqa: F401
+    FakeGitea,
+    app_with_fake_gitea,
+    provision_user_row,
+    tmp_env,
+)
+
+
+def _cfg():
+    from app.config import load_config
+    return load_config()
+
+
+def test_oauth_links_onto_existing_otc_email_row(app_with_fake_gitea):
+    from fastapi.testclient import TestClient
+    from app import auth, db
+
+    app, _fake = app_with_fake_gitea
+    with TestClient(app):
+        # An OTC-first user: email-only row, gitea_id NULL, still 'pending'.
+        db.conn().execute(
+            """
+            INSERT INTO users (gitea_id, gitea_login, email, display_name, avatar_url, role, permission_state)
+            VALUES (NULL, NULL, ?, ?, '', 'contributor', 'pending')
+            """,
+            ("dual@example.com", "dual"),
+        )
+        otc_id = db.conn().execute(
+            "SELECT id FROM users WHERE email = ? COLLATE NOCASE", ("dual@example.com",)
+        ).fetchone()["id"]
+
+        # Now the same human signs in via Gitea OAuth (new gitea_id, same email).
+        # Case-different email proves the NOCASE match.
+        user = auth.provision_user(
+            _cfg(),
+            {"id": 9001, "login": "dualgitea", "email": "Dual@Example.com",
+             "full_name": "Dual User", "avatar_url": "http://x/a.png"},
+        )
+
+        # Linked onto the SAME row — no second row, no 500.
+        assert user.user_id == otc_id
+        rows = db.conn().execute(
+            "SELECT id, gitea_id, gitea_login, permission_state, role FROM users WHERE email = ? COLLATE NOCASE",
+            ("dual@example.com",),
+        ).fetchall()
+        assert len(rows) == 1
+        assert rows[0]["id"] == otc_id
+        assert rows[0]["gitea_id"] == 9001           # OAuth identity attached
+        assert rows[0]["gitea_login"] == "dualgitea"
+        assert rows[0]["permission_state"] == "pending"  # admission state preserved
+        assert rows[0]["role"] == "contributor"
+
+
+def test_oauth_provisions_fresh_user_when_email_matches_no_one(app_with_fake_gitea):
+    from fastapi.testclient import TestClient
+    from app import auth, db
+
+    app, _fake = app_with_fake_gitea
+    with TestClient(app):
+        user = auth.provision_user(
+            _cfg(),
+            {"id": 9100, "login": "freshoauth", "email": "fresh@example.com",
+             "full_name": "Fresh", "avatar_url": ""},
+        )
+        row = db.conn().execute(
+            "SELECT gitea_id, gitea_login, role, permission_state FROM users WHERE id = ?",
+            (user.user_id,),
+        ).fetchone()
+        assert row["gitea_id"] == 9100
+        assert row["gitea_login"] == "freshoauth"
+        # Reaching provision_user means admission passed → granted (unchanged).
+        assert row["permission_state"] == "granted"
+
+
+def test_returning_oauth_user_matched_by_gitea_id_not_duplicated(app_with_fake_gitea):
+    from fastapi.testclient import TestClient
+    from app import auth, db
+
+    app, _fake = app_with_fake_gitea
+    with TestClient(app):
+        provision_user_row(user_id=55, login="returning", role="contributor")
+        before = db.conn().execute("SELECT COUNT(*) AS n FROM users").fetchone()["n"]
+
+        user = auth.provision_user(
+            _cfg(),
+            {"id": 55, "login": "returning-renamed", "email": "returning@test",
+             "full_name": "Returning", "avatar_url": ""},
+        )
+        after = db.conn().execute("SELECT COUNT(*) AS n FROM users").fetchone()["n"]
+
+        assert user.user_id == 55
+        assert after == before  # matched by gitea_id; no new row
+        row = db.conn().execute(
+            "SELECT gitea_login FROM users WHERE id = ?", (55,)
+        ).fetchone()
+        assert row["gitea_login"] == "returning-renamed"  # profile refreshed

backend/tests/test_prune_absent_projects.py

@@ -0,0 +1,163 @@
+"""§9 hardening — the registry reconcile prunes projects the registry no longer
+declares (was additive-only).
+
+A re-pin to a new registry used to strand the old projects' rows (PPE's
+stale-ecomm 1238-row entry cache starving writes). `prune_absent_projects` now
+deletes a removed project together with its collections and every project-scoped
+row, at whole-project granularity, with a `foreign_key_check` backstop. The
+safety guard is structural: `refresh_registry` raises on a read/transport error
+and `parse_registry` rejects an empty project list, so the prune is reached only
+after a real, non-empty registry parse — a transient failure can't wipe data.
+"""
+from __future__ import annotations
+
+import tempfile
+from pathlib import Path
+
+import pytest
+
+import app.db as db
+from app import projects, registry
+
+
+class _Cfg:
+    def __init__(self, path, default_id):
+        self.database_path = path
+        self.default_project_id = default_id
+
+
+def _two_project_db(monkeypatch, default_id="ohm"):
+    """A multi-project DB: default project 'ohm' (collection 'default') + a second
+    project 'ecomm' (collection 'ecomm'), each with a rich set of project- and
+    collection-keyed rows, plus the non-keyed `thread_messages` descendant."""
+    path = str(Path(tempfile.mkdtemp()) / "t.db")
+    cfg = _Cfg(path, default_id)
+    db.run_migrations(cfg)
+    monkeypatch.setattr(db, "_CONN", db.connect(path))
+    conn = db.conn()
+    conn.execute("DELETE FROM collections")
+    conn.execute("DELETE FROM projects")
+    conn.execute("INSERT INTO projects (id,name,content_repo,visibility) VALUES ('ohm','OHM','ohm-content','public')")
+    conn.execute("INSERT INTO projects (id,name,content_repo,visibility) VALUES ('ecomm','Ecomm','ecomm-content','public')")
+    conn.execute("INSERT INTO collections (id,project_id,type,subfolder,initial_state,visibility,name) "
+                 "VALUES ('default','ohm','document','','super-draft','public','OHM')")
+    conn.execute("INSERT INTO collections (id,project_id,type,subfolder,initial_state,visibility,name) "
+                 "VALUES ('ecomm','ecomm','bdd','ecomm','super-draft','public','Ecomm')")
+    conn.execute("INSERT INTO users (id,gitea_login,display_name,role) VALUES (1,'a','A','contributor')")
+
+    def seed(proj, coll, slug):
+        conn.execute("INSERT INTO cached_rfcs (slug,title,state,collection_id) VALUES (?,?,'active',?)",
+                     (slug, slug.title(), coll))
+        conn.execute("INSERT INTO cached_branches (rfc_slug,branch_name,collection_id) VALUES (?, 'main', ?)",
+                     (slug, coll))
+        conn.execute("INSERT INTO cached_prs (rfc_slug,pr_kind,repo,pr_number,title,state,project_id) "
+                     "VALUES (?, 'rfc_branch', ?, 1, 't', 'open', ?)", (slug, proj, proj))
+        conn.execute("INSERT INTO stars (user_id,rfc_slug,collection_id) VALUES (1,?,?)", (slug, coll))
+        conn.execute("INSERT INTO watches (user_id,rfc_slug,state,set_by,collection_id) VALUES (1,?,'watching','explicit',?)",
+                     (slug, coll))
+        conn.execute("INSERT INTO changes (rfc_slug,branch_name,kind,original,proposed,project_id) "
+                     "VALUES (?, 'main', 'ai', 'o', 'p', ?)", (slug, proj))
+        conn.execute("INSERT INTO notifications (recipient_user_id,event_kind,project_id) VALUES (1,'x',?)", (proj,))
+        cur = conn.execute("INSERT INTO threads (rfc_slug,anchor_kind,thread_kind,project_id) "
+                           "VALUES (?, 'whole-doc', 'chat', ?)", (slug, proj))
+        conn.execute("INSERT INTO thread_messages (thread_id,role,text) VALUES (?, 'user', 'hi')", (cur.lastrowid,))
+
+    seed("ohm", "default", "human")
+    seed("ecomm", "ecomm", "cart")
+    assert conn.execute("PRAGMA foreign_key_check").fetchall() == []  # seed is FK-clean
+    return cfg, conn
+
+
+_PRUNED_TABLES = [
+    "cached_rfcs", "cached_branches", "cached_prs", "stars", "watches",
+    "changes", "notifications", "threads",
+]
+
+
+def test_prune_removes_absent_project_and_all_its_data(monkeypatch):
+    cfg, conn = _two_project_db(monkeypatch, default_id="ohm")
+    pruned = projects.prune_absent_projects(cfg, {"ohm"})  # ecomm absent
+    assert pruned == ["ecomm"]
+
+    # ecomm: project, collection, and every keyed row gone.
+    assert conn.execute("SELECT 1 FROM projects WHERE id='ecomm'").fetchone() is None
+    assert conn.execute("SELECT 1 FROM collections WHERE id='ecomm'").fetchone() is None
+    for t in _PRUNED_TABLES:
+        col = "collection_id" if t in ("cached_rfcs", "cached_branches", "stars", "watches") else None
+        if col:
+            assert conn.execute(f"SELECT COUNT(*) n FROM {t} WHERE collection_id='ecomm'").fetchone()["n"] == 0, t
+        else:
+            assert conn.execute(f"SELECT COUNT(*) n FROM {t} WHERE project_id='ecomm'").fetchone()["n"] == 0, t
+    # The non-keyed descendant (thread_messages on ecomm's thread) is gone too.
+    assert conn.execute("SELECT COUNT(*) n FROM thread_messages").fetchone()["n"] == 1  # only ohm's remains
+
+    # ohm (the default) is fully intact.
+    assert conn.execute("SELECT 1 FROM projects WHERE id='ohm'").fetchone() is not None
+    assert conn.execute("SELECT 1 FROM collections WHERE id='default'").fetchone() is not None
+    assert conn.execute("SELECT COUNT(*) n FROM cached_rfcs WHERE slug='human'").fetchone()["n"] == 1
+    assert conn.execute("SELECT COUNT(*) n FROM threads WHERE rfc_slug='human'").fetchone()["n"] == 1
+
+    # FK integrity intact after the prune.
+    assert conn.execute("PRAGMA foreign_key_check").fetchall() == []
+
+
+def test_prune_is_noop_when_all_projects_present(monkeypatch):
+    cfg, conn = _two_project_db(monkeypatch, default_id="ohm")
+    assert projects.prune_absent_projects(cfg, {"ohm", "ecomm"}) == []
+    assert conn.execute("SELECT COUNT(*) n FROM projects").fetchone()["n"] == 2
+    assert conn.execute("SELECT COUNT(*) n FROM cached_rfcs").fetchone()["n"] == 2
+
+
+def test_prune_never_removes_the_default_project(monkeypatch):
+    # Even with an EMPTY present set (which the real caller can't produce —
+    # parse_registry rejects an empty registry), the default project survives.
+    cfg, conn = _two_project_db(monkeypatch, default_id="ohm")
+    pruned = projects.prune_absent_projects(cfg, set())
+    assert "ohm" not in pruned and "ecomm" in pruned
+    assert conn.execute("SELECT 1 FROM projects WHERE id='ohm'").fetchone() is not None
+    assert conn.execute("SELECT 1 FROM collections WHERE id='default'").fetchone() is not None
+    assert conn.execute("PRAGMA foreign_key_check").fetchall() == []
+
+
+def test_empty_registry_is_rejected_so_prune_is_never_reached():
+    # The structural safety guard: a parse that would yield no projects raises,
+    # so refresh_registry never reaches the prune with an empty present set.
+    with pytest.raises(registry.RegistryError):
+        registry.parse_registry("deployment:\n  name: x\nprojects: []\n")
+
+
+def test_incidental_refresh_does_not_prune(app_with_fake_gitea):  # noqa: F811
+    """An incidental re-mirror (the registry webhook, prune=False) must NOT prune
+    a project absent from projects.yaml — only the full-reconcile triggers
+    (startup + periodic sweep) prune. Guards the scoping decision so an in-app
+    refresh can't wipe a project."""
+    import hashlib
+    import hmac
+    import json as _json
+
+    from fastapi.testclient import TestClient
+    from app import db
+
+    app, _fake = app_with_fake_gitea
+    with TestClient(app) as client:
+        # A project present in the DB but not in the fake registry's projects.yaml.
+        db.conn().execute(
+            "INSERT INTO projects (id,name,content_repo,visibility) "
+            "VALUES ('ghost','Ghost','ghost-content','public')"
+        )
+        body = _json.dumps({"repository": {"full_name": "wiggleverse/registry"}}).encode()
+        secret = "test-webhook-secret-for-signature-verification"
+        sig = hmac.new(secret.encode(), body, hashlib.sha256).hexdigest()
+        r = client.post(
+            "/api/webhooks/gitea",
+            content=body,
+            headers={"X-Gitea-Event": "push", "X-Gitea-Signature": sig,
+                     "Content-Type": "application/json"},
+        )
+        assert r.status_code == 200
+        # The incidental re-mirror left the ghost in place (prune is off here).
+        assert db.conn().execute("SELECT 1 FROM projects WHERE id='ghost'").fetchone() is not None
+
+
+# app_with_fake_gitea fixture import for the integration test above.
+from test_propose_vertical import app_with_fake_gitea, tmp_env  # noqa: E402,F401

e2e/main-view-ask.spec.js

@@ -0,0 +1,53 @@
+import { test, expect } from './lib/fixtures.js'
+import { signIn, OWNER_EMAIL } from './lib/auth.js'
+import { dismissCookies } from './lib/ui.js'
+
+// §8.12 Option B — asking-while-reading on the canonical `main` view.
+//
+// Before this fix, the AI "Ask" affordance had no chat surface on main (the
+// human-discussion panel renders there, not the chat panel), so asking from the
+// reading view silently did nothing. Option B transparently cuts an edit branch,
+// navigates to it, and runs the question there.
+//
+// This spec asserts the model-independent core: the branch cut + navigation
+// (old code did NOTHING here — the silent no-op), and that the question turn
+// actually FIRED against the branch. The turn's outcome is environment-specific
+// — PPE/prod (a model is configured) streams an answer and the optimistic
+// question persists; the Tier-1 stack has no AI provider, so the turn errors
+// with "Chat failed". Either outcome proves the turn fired; we match both so the
+// one spec passes in both environments. The quote-passthrough + answer rendering
+// are covered deterministically by the RFCView unit tests (model mocked).
+//
+// Driven through the prompt bar, which shares handlePrompt → handleMainAsk with
+// the selection tooltip's onAsk. Runs against the faceted `bdd` collection entry
+// (collection-scoped — the same three-tier path G-15 hardened).
+const ENTRY = '/p/ohm/c/bdd/e/checkout-returning'
+
+test('Ask from the canonical view cuts an edit branch and lands the question in chat', async ({ page }) => {
+  await signIn(page, OWNER_EMAIL)
+  await page.goto(ENTRY)
+  await dismissCookies(page)
+
+  // We start on the canonical (main) view: read-only, no chat surface — the
+  // discuss-mode banner is the canonical-view tell, and the URL has no branch.
+  await expect(page.locator('.discuss-mode-banner')).toContainText('read-only')
+  expect(new URL(page.url()).searchParams.get('branch')).toBeNull()
+
+  // Ask a question from the prompt bar.
+  const question = `What problem does this entry solve? (${Date.now()})`
+  await page.locator('.prompt-input').fill(question)
+  await page.getByRole('button', { name: 'Ask', exact: true }).click()
+
+  // Option B: an edit branch is cut and we navigate onto it (no silent no-op).
+  await expect(async () => {
+    expect(new URL(page.url()).searchParams.get('branch')).toMatch(/^edit/)
+  }).toPass({ timeout: 30_000 })
+
+  // The question turn fired against the new branch: either the optimistic
+  // question is showing (model answered / answering) or the turn surfaced a
+  // chat error (no AI provider in Tier-1). Both prove it ran — the pre-fix
+  // silent no-op showed neither.
+  await expect(
+    page.getByText(question).or(page.getByText(/Chat failed|No AI providers/i)),
+  ).toBeVisible({ timeout: 30_000 })
+})

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "rfc-app-frontend",
   "private": true,
-  "version": "0.53.0",
+  "version": "0.55.0",
   "type": "module",
   "scripts": {
     "dev": "vite",

frontend/src/components/RFCView.jsx

@@ -93,6 +93,15 @@ export default function RFCView({ viewer }) {
   // surface to expose.
   const editorRef = useRef(null)
   const originalSourceLinesRef = useRef([])
+  // §8.12 Option B — asking-while-reading on `main`. When Ask is invoked from
+  // the canonical view we transparently cut an edit branch, navigate to it, and
+  // run the question there once its view (and main_thread_id) loads. The pending
+  // question is stashed here and fired by the pending-ask effect below — we
+  // can't push optimistic chat messages before branchView.main_thread_id exists.
+  const pendingAskRef = useRef(null)   // { text, quote, branch } | null
+  // Live ref to submitChatTurn so the (intentionally narrow-dep) message-load
+  // effect can fire the stashed §8.12 main-view ask with the latest closure.
+  const submitChatTurnRef = useRef(null)
   const [editorContent, setEditorContent] = useState('')
   // Mirror of the live CM6 doc for the Contribute-mode preview pane,
   // debounced so the preview doesn't re-render on every keystroke.
@@ -237,8 +246,21 @@ export default function RFCView({ viewer }) {
   // Load chat messages whenever the branch's main thread id resolves.
   useEffect(() => {
     if (!branchView?.main_thread_id) return
-    loadAllMessages(slug, branchParam, branchView.threads).then(setMessages)
+    loadAllMessages(slug, branchParam, branchView.threads).then(loaded => {
-  }, [branchView?.main_thread_id, slug, branchParam])
+      setMessages(loaded)
+      // §8.12 Option B — run the stashed main-view question now that the freshly
+      // cut branch's messages have loaded. Firing here (rather than in a parallel
+      // effect) appends the turn AFTER the loaded set, so this late message load
+      // can't clobber the optimistic chat turn. Land in contribute mode to
+      // surface any edits the turn proposes (AI chat is an editing activity).
+      const pending = pendingAskRef.current
+      if (pending && pending.branch === branchView.branch_name && pending.branch === branchParam) {
+        pendingAskRef.current = null
+        setMode('contribute')
+        submitChatTurnRef.current?.(pending.text, pending.quote)
+      }
+    })
+  }, [branchView?.main_thread_id, branchView?.branch_name, slug, branchParam])
 
   // Selection wiring (§8.12). MarkdownPreview reports {text, coords}
   // sourced from window.getSelection(); the SelectionTooltip
@@ -335,8 +357,12 @@ export default function RFCView({ viewer }) {
   }, [manualCountdown, flushManualBuffer])
 
   // ── Start contributing ─────────────────────────────────────────────────
+  // Returns the branch the viewer should now be on (the freshly cut branch on
+  // main, or the current branch on a mode-flip), or null if no branch was cut
+  // (signed-out → login redirect, or the server rejected the cut). §8.12's
+  // main-view Ask reuses this dispatch and consumes the returned branch name.
   const handleStartContributing = useCallback(async () => {
-    if (!viewer) { window.location.href = '/auth/login'; return }
+    if (!viewer) { window.location.href = '/auth/login'; return null }
     if (branchParam === 'main') {
       try {
         // §9.5 dispatch: super-drafts cut a meta-repo edit branch via
@@ -345,18 +371,34 @@ export default function RFCView({ viewer }) {
           ? await startEditBranch(slug)
           : await promoteToBranch(slug)
         setSearchParams({ branch: branch_name })
+        return branch_name
       } catch (err) {
         setError(err.message)
+        return null
       }
-      return
     }
     // Non-main: pure mode flip per §8.14.
     if (pendingDiscussChanges.length > 0) {
       setPendingDiscussChanges([])
     }
     setMode('contribute')
+    return branchParam
   }, [viewer, slug, branchParam, pendingDiscussChanges, setSearchParams, isSuperDraft])
 
+  // ── §8.12 Option B — Ask invoked from the canonical `main` view ─────────
+  // AI chat is an editing activity and only runs on an edit branch, so asking
+  // while reading transparently cuts one (same dispatch as Start Contributing),
+  // navigates to it, and stashes the question; the pending-ask effect fires it
+  // once the new branch's view (and its main_thread_id) has resolved. Degrades
+  // gracefully: signed-out → login; a viewer who can't contribute has the cut
+  // rejected server-side → the error surfaces and we neither navigate nor chat.
+  const handleMainAsk = useCallback(async (text, quote) => {
+    if (!viewer) { window.location.href = '/auth/login'; return }
+    const branch = await handleStartContributing()
+    if (!branch || branch === 'main') return   // cut failed → setError already ran
+    pendingAskRef.current = { text, quote: quote || null, branch }
+  }, [viewer, handleStartContributing])
+
   // ── Submit a chat turn (prompt bar or selection tooltip) ───────────────
   const submitChatTurn = useCallback(async (text, quote) => {
     if (!branchView?.main_thread_id || isStreaming) return
@@ -427,16 +469,24 @@ export default function RFCView({ viewer }) {
     }
   }, [slug, branchParam, branchView?.main_thread_id, isStreaming, viewer, selectedModel, mode])
 
+  // Keep submitChatTurnRef pointed at the latest submitChatTurn so the
+  // message-load effect's §8.12 main-view ask uses the current branch closure.
+  useEffect(() => { submitChatTurnRef.current = submitChatTurn }, [submitChatTurn])
+
   const handlePrompt = useCallback((text, sel) => {
     const quote = sel?.text || null
+    // On the canonical view there is no chat surface; Ask cuts an edit branch
+    // and runs the question there (§8.12 Option B). On a branch, ask directly.
+    if (branchParam === 'main') { handleMainAsk(text, quote); return }
     submitChatTurn(text, quote)
-  }, [submitChatTurn])
+  }, [branchParam, handleMainAsk, submitChatTurn])
 
   const handleTooltipAsk = useCallback(async (textOrNull, quote) => {
     if (textOrNull === null) { setSelection(null); return }
     setSelection(null)
+    if (branchParam === 'main') { await handleMainAsk(textOrNull, quote); return }
     await submitChatTurn(textOrNull, quote)
-  }, [submitChatTurn])
+  }, [branchParam, handleMainAsk, submitChatTurn])
 
   const handleTooltipFlag = useCallback(async (label, quote) => {
     if (!viewer) { window.location.href = '/auth/login'; return }

frontend/src/components/RFCView.test.jsx

@@ -0,0 +1,220 @@
+import React from 'react'
+import { describe, it, expect, vi, beforeEach } from 'vitest'
+import { render, screen, fireEvent, waitFor } from '@testing-library/react'
+import { MemoryRouter, Routes, Route } from 'react-router-dom'
+
+// §8.12 Option B — asking-while-reading on the canonical `main` view. The AI
+// "Ask" affordance (selection tooltip + prompt bar) has no chat surface on
+// main (RFCDiscussionPanel renders there, not ChatPanel). These tests cover the
+// transparent branch-cut: Ask on main cuts an edit branch, navigates to it, and
+// runs the question there; Ask on a branch is unchanged; a rejected cut degrades
+// gracefully; Flag on main still opens a discussion thread.
+
+const getRFC = vi.fn()
+const getRFCMain = vi.fn()
+const getBranch = vi.fn()
+const listModels = vi.fn()
+const getCollection = vi.fn()
+const getThreadMessages = vi.fn()
+const promoteToBranch = vi.fn()
+const startEditBranch = vi.fn()
+const streamChatTurn = vi.fn()
+const createThread = vi.fn()
+
+vi.mock('../api', () => ({
+  getRFC: (...a) => getRFC(...a),
+  getRFCMain: (...a) => getRFCMain(...a),
+  getBranch: (...a) => getBranch(...a),
+  listModels: (...a) => listModels(...a),
+  getCollection: (...a) => getCollection(...a),
+  getThreadMessages: (...a) => getThreadMessages(...a),
+  promoteToBranch: (...a) => promoteToBranch(...a),
+  startEditBranch: (...a) => startEditBranch(...a),
+  streamChatTurn: (...a) => streamChatTurn(...a),
+  createThread: (...a) => createThread(...a),
+  acceptChange: vi.fn(), declineChange: vi.fn(), editMetadata: vi.fn(),
+  manualFlush: vi.fn(), reaskChange: vi.fn(), resolveThread: vi.fn(),
+  setBranchVisibility: vi.fn(), claimOwnership: vi.fn(),
+  retireRFC: vi.fn(), unretireRFC: vi.fn(),
+}))
+
+vi.mock('../lib/entryPaths', () => ({
+  entryPath: () => '/p/ohm/c/rfc-app/e/the-entry',
+  entryPrPath: () => '/pr',
+  useProjectId: () => 'ohm',
+  useCollectionId: () => 'rfc-app',
+}))
+vi.mock('../lib/analytics', () => ({ EVENTS: { RFC_VIEWED: 'rfc_viewed' }, track: vi.fn() }))
+
+// Stub the heavy children; expose just the callbacks/data the flow needs.
+vi.mock('./MarkdownSourceEditor.jsx', () => ({ default: () => null }))
+vi.mock('./MarkdownPreview.jsx', () => ({ default: () => <div data-testid="preview" /> }))
+vi.mock('./RFCDiscussionPanel.jsx', () => ({ default: () => <div data-testid="discussion-panel" /> }))
+vi.mock('./ChatPanel.jsx', () => ({
+  default: ({ messages }) => (
+    <div data-testid="chat-panel">
+      {(messages || []).map((m, i) => (
+        <div key={i} data-role={m.role}>{m.text}</div>
+      ))}
+    </div>
+  ),
+}))
+vi.mock('./SelectionTooltip.jsx', () => ({
+  default: ({ onAsk, onFlag, disabled }) => (
+    <div data-testid="tooltip">
+      <button data-testid="tooltip-ask" disabled={disabled}
+        onClick={() => onAsk('Why this approach?', 'the selected quote')}>ask</button>
+      <button data-testid="tooltip-flag"
+        onClick={() => onFlag('confusing', 'the selected quote')}>flag</button>
+    </div>
+  ),
+}))
+vi.mock('./PromptBar.jsx', () => ({
+  default: ({ onSubmit, disabled }) => (
+    <button data-testid="prompt-submit" disabled={disabled}
+      onClick={() => onSubmit('Why this approach?', { text: 'the selected quote' })}>send</button>
+  ),
+}))
+vi.mock('./ChangePanel.jsx', () => ({ default: () => null, diffWords: () => [] }))
+vi.mock('./PRModal.jsx', () => ({ default: () => null }))
+vi.mock('./GraduateDialog.jsx', () => ({ default: () => null }))
+vi.mock('./InvitationsModal.jsx', () => ({ default: () => null }))
+vi.mock('./MetadataFieldsPanel.jsx', () => ({ default: () => null }))
+
+import RFCView from './RFCView.jsx'
+
+const ACTIVE_ENTRY = {
+  id: 'RFC-0001', title: 'The Entry', state: 'active',
+  owners: [], meta: {}, proposed_use_case: '', tags: [],
+}
+const MAIN_VIEW = { slug: 'the-entry', branches: [], open_prs: [] }
+
+function branchPayload(branch, { main_thread_id = 't-1' } = {}) {
+  return {
+    slug: 'the-entry', title: 'The Entry', branch_name: branch, body: '# Body\n',
+    body_sha: 'abc', main_thread_id, threads: [], changes: [],
+    visibility: {}, grants: [], creator: 'me',
+    capabilities: { can_contribute: true, can_change_branch_settings: true },
+  }
+}
+
+const VIEWER = { gitea_login: 'me', role: 'contributor' }
+
+function renderView(viewer = VIEWER, initialBranch = null) {
+  const path = initialBranch ? `/e/the-entry?branch=${initialBranch}` : '/e/the-entry'
+  return render(
+    <MemoryRouter initialEntries={[path]}>
+      <Routes>
+        <Route path="/e/:slug" element={<RFCView viewer={viewer} />} />
+      </Routes>
+    </MemoryRouter>,
+  )
+}
+
+describe('RFCView — §8.12 main-view Ask (Option B)', () => {
+  beforeEach(() => {
+    vi.clearAllMocks()
+    getRFC.mockResolvedValue(ACTIVE_ENTRY)
+    getRFCMain.mockResolvedValue(MAIN_VIEW)
+    listModels.mockResolvedValue({ models: [{ id: 'claude' }], default: 'claude' })
+    getCollection.mockResolvedValue({ fields: null })
+    getThreadMessages.mockResolvedValue({ messages: [] })
+    getBranch.mockImplementation((_slug, branch) => Promise.resolve(branchPayload(branch)))
+    promoteToBranch.mockResolvedValue({ branch_name: 'edit-me-1' })
+    startEditBranch.mockResolvedValue({ branch_name: 'edit-the-entry' })
+    createThread.mockResolvedValue({ thread_id: 1, message_id: 1 })
+    streamChatTurn.mockImplementation(async (_slug, _branch, _threadId, { text, quote }, cb) => {
+      cb.onChunk?.('Here is the answer.')
+      cb.onChanges?.({ message_id: 42 })
+      cb.onDone?.()
+      return { assistantId: 42, userMsgId: 41 }
+    })
+  })
+
+  it('cuts an edit branch and runs the question (with the quote) there', async () => {
+    renderView()
+    await screen.findByTestId('discussion-panel')   // we start on main
+
+    fireEvent.click(screen.getByTestId('prompt-submit'))
+
+    // The active-RFC dispatch is promote-to-branch.
+    await waitFor(() => expect(promoteToBranch).toHaveBeenCalledWith('the-entry'))
+
+    // Once the new branch view loads, the turn runs there — not on main —
+    // with the selected quote intact.
+    await waitFor(() => expect(streamChatTurn).toHaveBeenCalled())
+    const [slug, branch, threadId, payload] = streamChatTurn.mock.calls[0]
+    expect(slug).toBe('the-entry')
+    expect(branch).toBe('edit-me-1')
+    expect(threadId).toBe('t-1')
+    expect(payload).toMatchObject({ text: 'Why this approach?', quote: 'the selected quote' })
+
+    // The chat panel is now mounted (we left main) and shows the streamed answer.
+    const panel = await screen.findByTestId('chat-panel')
+    await waitFor(() => expect(panel).toHaveTextContent('Here is the answer.'))
+  })
+
+  it('uses start-edit-branch for a super-draft entry', async () => {
+    getRFC.mockResolvedValue({ ...ACTIVE_ENTRY, state: 'super-draft' })
+    getBranch.mockImplementation((_slug, branch) =>
+      Promise.resolve(branchPayload(branch === 'main' ? 'main' : 'edit-the-entry')))
+    renderView()
+    await screen.findByTestId('discussion-panel')
+
+    fireEvent.click(screen.getByTestId('tooltip-ask'))
+
+    await waitFor(() => expect(startEditBranch).toHaveBeenCalledWith('the-entry'))
+    expect(promoteToBranch).not.toHaveBeenCalled()
+    await waitFor(() => expect(streamChatTurn).toHaveBeenCalled())
+    expect(streamChatTurn.mock.calls[0][1]).toBe('edit-the-entry')
+  })
+
+  it('asking from an existing branch runs directly, without cutting a new one', async () => {
+    renderView(VIEWER, 'edit-me-1')
+    await screen.findByTestId('chat-panel')
+
+    fireEvent.click(screen.getByTestId('prompt-submit'))
+
+    await waitFor(() => expect(streamChatTurn).toHaveBeenCalled())
+    expect(promoteToBranch).not.toHaveBeenCalled()
+    expect(startEditBranch).not.toHaveBeenCalled()
+    expect(streamChatTurn.mock.calls[0][1]).toBe('edit-me-1')
+  })
+
+  it('surfaces an error and does not chat when the branch cut is rejected', async () => {
+    promoteToBranch.mockRejectedValue(new Error('not a contributor'))
+    renderView()
+    await screen.findByTestId('discussion-panel')
+
+    fireEvent.click(screen.getByTestId('prompt-submit'))
+
+    await waitFor(() => expect(promoteToBranch).toHaveBeenCalled())
+    // Explicit error path — not a silent no-op — and no spurious chat turn.
+    expect(await screen.findByText(/not a contributor/)).toBeInTheDocument()
+    expect(streamChatTurn).not.toHaveBeenCalled()
+  })
+
+  it('Flag on the canonical view still creates a discussion thread (no branch cut)', async () => {
+    renderView()
+    await screen.findByTestId('discussion-panel')
+
+    fireEvent.click(screen.getByTestId('tooltip-flag'))
+
+    await waitFor(() => expect(createThread).toHaveBeenCalled())
+    const [, branch, body] = createThread.mock.calls[0]
+    expect(branch).toBe('main')
+    expect(body).toMatchObject({ thread_kind: 'flag' })
+    expect(promoteToBranch).not.toHaveBeenCalled()
+    expect(streamChatTurn).not.toHaveBeenCalled()
+  })
+
+  it('signed-out viewer gets a read-only path on main — no prompt bar, Ask disabled, no cut', async () => {
+    renderView(null)
+    await screen.findByTestId('discussion-panel')
+
+    expect(screen.queryByTestId('prompt-submit')).toBeNull()
+    expect(screen.getByTestId('tooltip-ask')).toBeDisabled()
+    expect(promoteToBranch).not.toHaveBeenCalled()
+    expect(startEditBranch).not.toHaveBeenCalled()
+  })
+})