diff --git a/GOVERNANCE.md b/GOVERNANCE.md
index 7bb7c42f..371fa97a 100644
--- a/GOVERNANCE.md
+++ b/GOVERNANCE.md
@@ -25,6 +25,7 @@ A state-changing GitHub call is the highest-blast-radius thing an agent does her
 - **Write only within the owner of the current project's repository.** Every state-changing call targets this project's `origin` or another repository under the same owner, which is the fleet the maintainer already administers. A broad or logged-in identity is capability, not permission: a token that *can* reach another owner's repository does not authorize writing to it. Writing under a **different owner** needs explicit human permission naming that repository, granted deliberately rather than assumed from a token's reach, and a "harmless test" write is still a write, so there is no probe exception. That boundary is where the harm sits, since the incident this rule exists for was a stray comment on a stranger's repository, not work across the maintainer's own projects. Reads from anywhere are fine.
 - **Never fabricate, guess, or reuse an identifier passed to a write.** Every id a state-changing call consumes (a node id, a numeric id, a thread or comment id) is captured from a live query in the **same** session into a variable and passed from there. Do not hand-type an id, guess it, recall it from memory or an earlier session, or copy it from documentation or an example. Ids commonly resolve **globally**, so a wrong-but-valid id does not fail. It writes to the wrong target, in someone else's repository. If a query returns no id, stop rather than invent one to proceed.
 - **A write is never a probe, and a write's output is never suppressed.** Never fire a state-changing call to see whether it works: decide it should happen, make it happen, and read the result. Never append output-discarding redirection or a force-success tail to a mutation (for example `>/dev/null`, `2>/dev/null`, `&>/dev/null`, `|| true`, `|| :`, `|| echo`), because the write's output is exactly what must be read. A write that appears to fail is **verified, not assumed harmless**, because the operation may have succeeded on the server while the client reported an error, so confirm the actual state before retrying or moving on. The ban targets hiding a *failure*. An ad-hoc call's response is the only signal you get, so `>/dev/null 2>&1`, `|| true`, and `|| echo`, which swallow the error stream or force success, are never acceptable on one. A committed script under `set -e` is a narrow exception: it may send a write's *stdout* to `/dev/null` to drop the success-response noise, because stderr stays visible and a failed write still aborts loudly (`repo-config/configure.sh` does exactly this). The exception is stdout-only suppression inside a reviewed, fail-loud script, never `2>&1` or a force-success tail, and never an ad-hoc command.
+- **A refused write is reported, never re-shaped, and the maintainer's say-so does not lift a refusal by the harness.** These are two different permissions and only one of them is the maintainer's to give. When the agent harness refuses a write, the maintainer authorizing it in conversation does not change the outcome, and the identical call is refused again, so a second attempt is not worth making and reading the second refusal as a flake is how an agent starts hunting for another shape of the same request. **That hunt is the failure this rule exists to stop.** Re-expressing a refused `gh` command as a raw `gh api -X POST` reaches the same endpoint with the same identity and the same blast radius, having defeated the one control that stopped it, and it is the more dangerous version because the agent believes it has permission. So a refused write is never re-attempted through a different API surface, a different tool, or a rephrasing, and it is never routed around by the agent writing itself a permission rule, which is self-authorization whatever the maintainer said. Two routes remain, both of them the maintainer's: they add the permission rule themselves, or they run the command themselves. Raise it as a blocked decision naming those two (see "Communicating with the User"), and where the work needs the result rather than the call, say what the agent will verify once the maintainer has run it. **A refusal is also a fact about the contract, not just about the session**: where a required verification can only be performed by a write the agent is refused, the document requiring it says so and names who runs it, since a check that is mandatory and unperformable is quietly dropped and then reported as done.
 
 ## Representative Data in Agent-Authored Text
 
diff --git a/TODO.md b/TODO.md
index c2369fa0..73e9d018 100644
--- a/TODO.md
+++ b/TODO.md
@@ -30,7 +30,7 @@ The gates in [`scripts/`][scripts] are hub-only and are the fleet's main defense
 
 Everything here changes a file the fleet holds a copy of, so each entry costs a hub edit plus a sweep, and the sweep is the part that gets skipped.
 
-- Re-vendor the changed `verbatim` content across the fleet, which is one sweep covering five files. `repo-config/configure.sh` is carried `verbatim` with `appliesTo: "*"` and the hub swept it to one sentence per line. In `AGENTS.md`, "Context and Delegation Discipline" carries the wait rule's failure clause and "Where the Rules Live" carries a row for the new section named below. In `GOVERNANCE.md`, "Verification Discipline" carries the rule that a launched process is not a result, "PR Review Etiquette" carries the five outcomes that close a finding, and "Representative Data in Agent-Authored Text" is an entirely new carried section that no downstream repo holds, which the audit reports as a missing section rather than as drift. Three further `GOVERNANCE.md` sections differ by a single word each, "Documentation Style Conventions", "Communicating with the User" and "Repository Details", where the format's name was capitalized to the convention `CODESTYLE.md` "Markdown and Spelling" now states, so they are byte-mismatched for a reason a reader of the diff would otherwise call cosmetic. Two comment lines in `.markdownlint-cli2.jsonc` took the same capitalization, and that file is `verbatim` and `whole`, so every downstream copy is byte-mismatched on a config nothing else changed about. `CODESTYLE.md` carries the new item and is the fifth file, at `intent` rather than `verbatim` fidelity, so it reaches the fleet as a rule each repo adopts in its own copy rather than as bytes to match, and the same mixed spelling is waiting in every downstream tree. Every repo already holding a copy of a changed section is byte-mismatched against the hub until it takes the new one, which the audit reports as stale rather than modified. This sweep is also the follow-through [#489][issue-489] and [#379][issue-379] were waiting on, and the `.editorconfig` line in [#353][issue-353] rides the same visit to each repo. Regenerate [reports/divergences.md][divergences-report] before using it as the work list, since the committed copy is dated 2026-07-22 and therefore predates the router split, which shows in it naming "Repository Boundaries and Write Safety", "Git and Commit Rules" and "Verification Discipline" as `AGENTS.md` sections when all three now live in `GOVERNANCE.md`. A stale ledger is the same hazard as a stale exemption, in that it hands out a work list measured against a tree that no longer exists.
+- Re-vendor the changed `verbatim` content across the fleet, which is one sweep covering five files. `repo-config/configure.sh` is carried `verbatim` with `appliesTo: "*"` and the hub swept it to one sentence per line. In `AGENTS.md`, "Context and Delegation Discipline" carries the wait rule's failure clause and "Where the Rules Live" carries a row for the new section named below. In `GOVERNANCE.md`, "Verification Discipline" carries the rule that a launched process is not a result, "PR Review Etiquette" carries the five outcomes that close a finding, "Repository Boundaries and Write Safety" carries the rule that a refused write is reported rather than re-shaped, and "Representative Data in Agent-Authored Text" is an entirely new carried section that no downstream repo holds, which the audit reports as a missing section rather than as drift. Three further `GOVERNANCE.md` sections differ by a single word each, "Documentation Style Conventions", "Communicating with the User" and "Repository Details", where the format's name was capitalized to the convention `CODESTYLE.md` "Markdown and Spelling" now states, so they are byte-mismatched for a reason a reader of the diff would otherwise call cosmetic. Two comment lines in `.markdownlint-cli2.jsonc` took the same capitalization, and that file is `verbatim` and `whole`, so every downstream copy is byte-mismatched on a config nothing else changed about. `CODESTYLE.md` carries the new item and is the fifth file, at `intent` rather than `verbatim` fidelity, so it reaches the fleet as a rule each repo adopts in its own copy rather than as bytes to match, and the same mixed spelling is waiting in every downstream tree. Every repo already holding a copy of a changed section is byte-mismatched against the hub until it takes the new one, which the audit reports as stale rather than modified. This sweep is also the follow-through [#489][issue-489] and [#379][issue-379] were waiting on, and the `.editorconfig` line in [#353][issue-353] rides the same visit to each repo. Regenerate [reports/divergences.md][divergences-report] before using it as the work list, since the committed copy is dated 2026-07-22 and therefore predates the router split, which shows in it naming "Repository Boundaries and Write Safety", "Git and Commit Rules" and "Verification Discipline" as `AGENTS.md` sections when all three now live in `GOVERNANCE.md`. A stale ledger is the same hazard as a stale exemption, in that it hands out a work list measured against a tree that no longer exists.
 - Drop the unused `GITHUB_TOKEN` grants from the merge-bot workflow, per [#521][issue-521], which the tree confirms is still open, since all four jobs still carry `contents: write` and `pull-requests: write` while every write in the file authenticates with the App token. The finding is least privilege on the highest-blast-radius shape in the fleet, a `pull_request_target` workflow holding an App private key, where the grant is not exploitable today only because no step consumes it. Three things to settle: whether to drop the job-level blocks or set a workflow-level `permissions: {}`, whether the audit should compare permissions at all given the file is `interface` fidelity with only a `requiredJobKeys` contract, and the re-vendor, since every repo carries the file. Note that [`spec/files.json`][files] now declares this workflow at `appliesTo: "*"`, which closes the separate gap [#456][issue-456] raised, that D8.1 graded a file the file spec never required.
 - State the rule that an agent never assumes a Docker image is present locally, however recently it pulled one, because a background prune can remove an image between two commands of the same session and the run then fails on something that worked minutes earlier. The prescription is the always-pull flag by default, and an explicit `docker pull <image>:<tag>` before use where the flag does not apply. The four documented lint invocations already carry `--pull=always`, in [`GOVERNANCE.md`][governance] "Running the Linters Locally", mirrored in [`OPERATIONS.md`][operations] and both VS Code task snippets, so what is missing is the rule rather than the one-liners: an agent composing an ad-hoc `docker run` drops the flag precisely because it believes the image is already cached. Two things to settle when writing it. Where it lives, since that section is scoped to the four lint tools while the rule covers any container an agent starts, which argues for placing it with the Docker or verification guidance instead, and whether it is carried to the fleet, since every repo runs the same images from the same instructions. Note the limit before writing it, so the flag does not read as the whole answer: `docker run` against a registry tag re-pulls an absent image on its own, so the cases that actually break are a locally built tag with no registry to pull from, and any command that branches on the image being present (`docker image inspect`, `docker images`) rather than simply running it.
 - Decide where a carried file may name hub-only machinery, since `GOVERNANCE.md` "PR Review Etiquette" points at `scripts/pr_review.py` and the fleet carries the section but not the script. A downstream reader follows that pointer to a path their repo does not have. Either the script joins the carried set, or the rule states the behavior and drops the tool name the way the coordination-reference rule already requires for the template repo itself.
diff --git a/WORKFLOW.md b/WORKFLOW.md
index de19870e..dcc1e97e 100644
--- a/WORKFLOW.md
+++ b/WORKFLOW.md
@@ -263,6 +263,7 @@ For each *applicable* scenario, evaluate every job's `if:`/`needs:` against the
 - Drive a `smoke: true` push-probe of the build task for **both** the default and a non-default branch and assert the version classification (clean vs prerelease) and that the gate passes, **without publishing**. *Caveat: the Docker leg logs in to the registry even on smoke and reads the buildcache, so it needs `DOCKER_HUB_*` secrets and cannot run on a fork PR (same-repo only).*
 - Per registry: after a real publish, query NuGet.org for the expected version + prerelease classification (and the `.snupkg` on the symbol server), and confirm a re-run added no duplicate. For PyPI inspect the `Compute PyPI version step` log and the built `dist/*` filenames for `.dev0` off `develop` vs a plain version on the default branch.
 - Inspect the latest real publish's logs for `PublicRelease`/`SemVer2` per leg and confirm the artifact lifecycle (uploaded, consumed, deleted, with none left behind).
+- **The deploy ref gate (S13) is verified only by tripping it, and the dispatch that trips it is the maintainer's to run.** Dispatch the production environment from a non-default ref and expect the run to fail at the gate. The evidence is four things, and each of them matters: the gate job's conclusion, its error text naming the expected and the received ref, every downstream job recorded as **skipped** rather than passed, and the deployment count against the production environment unchanged. Capture all four, because a gate that fails open and a gate nobody tripped produce the same empty run history, so "we have never seen it fail" is not evidence about the one control standing between a mis-dispatch and the live site. **The agent prepares the command and reads all four back afterwards. It does not fire it.** An agent harness may refuse to dispatch a production deploy, which is the harness working as intended, and the refusal is neither re-shaped into a raw API call nor talked around (GOVERNANCE.md "Repository Boundaries and Write Safety"). The same split applies to any probe that acts on the deploy host directly, an outbound SSH exercising a forced command among them.
 
 ### Assessment