# Contract — P13-S16: the projection gets maintained

**Status:** **DRAFT — UNBLOCKED 2026-08-09. NOT RATIFIED, and therefore NOT
dispatchable.**

**P13-S27 landed and was accepted at `4df8e25`**, so pin 0's blocker is discharged: an
authority now defines the implementation's current reduction semantics
(`epiphany_ops::CURRENT_REDUCTION_ALGORITHM_VERSION`, currently **`0`**), and
`core_spec.tex:11614`'s requirement is **met** — a stale canonical base is refused with
`CanonicalBaseRequiresRebuild` on both the read and write paths.

**What remains before dispatch is this contract's own ratification.** It is complete and
ratifiable as a *plan* and has **not been through adversarial review rounds**.
*Unblocked* and *dispatchable* are different states, and conflating them is how the S27
contract came to be ratified after a single round and then have that ratification
withdrawn — see its status block.

**This rung's first act is bumping the authority to `1`**, because it changes
`CreateStaffGroup`'s reduction verdict. That bump is the discipline S27 installed: no
mechanism can detect a semantics change, so the bump is the entire guarantee.

> **The original status, retained:** *DRAFT — BLOCKED on P13-S27. Not executable as
> written. Pin 0 exposes that no authority defines the implementation's current
> reduction semantics, and prose saying old canonical bases "must be rebuilt" does not
> make them unusable. `core_spec.tex:11614`'s requirement stays unmet until S27 supplies
> a version authority and a rejection-or-rebuild path.*

**Rung type:** **canonical reduction-semantics change.** This is stronger than
"behaviour change" and the first draft of this contract understated it. The same
operation set now reduces to a different canonical `Score`: a graph that reduces
cleanly today is refused, and a field that was never written is written.
Disposition **A** of `spec/CONTRACT_GENESIS_G3A_ENTITIES.md` §1.1, ratified
there as "the later maintenance/enforcement fix" and sequenced after G3b — which
has landed.

**No schema major or minor moves and no wire byte changes** — but the Operation
Catalog's version *does* move (pin 10), and the reduction-semantics question is
**separate from the schema/wire version question**. See pin 0.

**No policy ruling is owed.** Disposition A is already ratified. What follows is
its implementation.

---

## §0. What was verified before drafting

Read out of the working tree at `f876836`, not recalled. Every line number below
was confirmed by reading the line.

### 0.1 The refusal needs no new machinery — and this is the rung's biggest saving

Disposition A says `CreateStaffGroup` "MUST carry `members: []`; a non-empty
`members` is refused or normalized away at construction." An earlier reading of
this contract assumed refusal required a new `PreconditionFailureReason` — which
would have made S16 a schema-minor epoch event with a footprint in
`binary_format.tex`, `PLAN_GMINOR_SCHEMA_MINOR.md`, `decode.rs`, and the history
test. **It does not.**

`reduce.rs:1236` already defines the shared helper, and its own doc comment
already covers this case:

> The precondition no-op a structural create or delete returns when a container
> is non-empty where the operation requires it empty (**a create carrying
> children**, or a delete of a container with live children).

Three creates already call it for exactly this reason — `create_region`
(`:4174`), `create_staff_instance` (`:4246`), `create_voice` (`:4310`) — each
refusing a carried value that bears a separately-minted typed child.
`create_staff_group` is **the sole outlier**: it accepts carried `members` and
merely checks each is live (`:4489`–`:4502`).

So the refusal is the **fourth instance of an established pattern**, not a new
one. `PreconditionFailureReason::ContainerNotEmpty` stays at discriminant 10.
**No new reason, no epoch, no wire change, no accept-set move.**

### 0.2 Normalizing at construction or decode is excluded, on ratified grounds

Disposition A's "or normalized away" alternative must **not** be taken. The
committed decode vector at `ops/src/vectors.rs:829` pins 130 literal bytes of a
`CreateStaffGroup` envelope whose payload carries `members = [StaffId(…)]`, and
its contract is decode-then-re-encode injectivity. The text-projection golden at
`textproj/src/vectors.rs:198`–`:201` carries the same value. Folding `members`
away at decode would break both, and violates `req:binfmt:decode-vectors`
(`binary_format.tex:3404`ff): *"Canonical decode is injective: distinct byte
strings denote distinct values."*

**Refusal happens at reduction. Decode is untouched, and both artifacts survive
unchanged.** This is the same fork P13-S8 faces, resolved the same way, for the
same ratified reason.

### 0.3 The authorship cycle makes empty-at-mint the only self-consistent rule

`create_staff` requires a carried `group` to be live (`:4372`–`:4383`);
`create_staff_group` requires each carried member to be live (`:4489`–`:4502`).
There is **no `ModifyStaff`, `ModifyStaffGroup`, `DeleteStaff`, or
`DeleteStaffGroup`** anywhere in `OperationKind` (`payload.rs:170`–`:307`;
confirmed normatively at `operation_catalog.tex:1288` and `:1553`). With mints
only, no authoring order produces an agreeing pair.

Therefore the *only* authorable agreeing sequence is: mint the group empty, then
mint staves naming it. Requiring `members: []` is not an arbitrary restriction —
it is the sole rule consistent with the operation surface that exists.

### 0.4 The re-carry hazard, and the base-ingest hazard hiding behind it

`create_staff_group`'s idempotence check compares `op.group` against
`self.staff_group_values` (`:4466`–`:4469`). If maintenance wrote the appended
members into that map, a byte-identical re-carry of `CreateStaffGroup(g, [])`
would compare `[]` against `[s]` and return `RecreateContentMismatch` instead of
`AlreadyApplied`. Disposition A names this sub-pin explicitly.

**The non-obvious half:** base ingest (`:1611`–`:1620`) reseeds
`staff_group_values` from `score.staff_groups` — the *maintained* value. So even
if reduction keeps carried and derived apart in one session, a snapshot round
trip launders the derived value into the carried slot, and the same misverdict
returns **after a reload**. A test that never reloads cannot see it.

The resolution falls out of §0.1: once a non-empty carried `members` is refused,
**the carried value is by construction always empty**, so the base seed can
reconstruct it exactly rather than approximately (pin 4).

### 0.5 `t8b` is the obstacle, and it inverts rather than dies

`t8b_both_permitted_stale_forms_hold` (`reduce.rs:16339`, doc `:16317`–`:16337`)
asserts both stale forms **hold**, and its doc block names disposition A's two
maintenance rules as **mutations that must make those assertions fail**. The
test is a correctly-built detector for precisely this change.

It is therefore rewritten, not deleted: the same two authoring orders, with the
verdicts inverted — the missing order now yields agreement, the spurious order
now yields `ContainerNotEmpty`. Its doc block's mutation notes become the
rung's own mutation evidence, pointing the other way.

### 0.6 Invariant 21 has real work left after maintenance

Maintenance plus refusal does not make disagreement impossible:

- **Undo.** `reduce.rs:2967` removes a `Staff` from `score.staves` on undo but
  leaves its id in any live group's `members`. The reverse direction *is*
  guarded (`:6736`–`:6744`, which blocks undoing a group still named by a live
  staff); **this direction is not.**
- **Base ingest.** A blob authored before this rung, or by another
  implementation, can carry a disagreeing pair straight in.

There are 20 invariants (`invariants.rs:149`, count guard `:6064`). 21 is free.

---

## §1. Pins

**Pin 0 — declare the reduction-semantics break; do not pretend it is
containable.**

`core_spec.tex:14369`–`:14372` is normative: *"Two replicas with different
`ReductionAlgorithmVersion` may produce different canonical states from the same
operation set; the active superblock declares the version under which the
bundle's canonical base was materialized."* And `:11614`–`:11617`: *"Snapshots
produced under an earlier algorithm version cannot be used as canonical bases
under a later one without rebuilding."* This rung is exactly such a change.

**The obvious disposition — bump the version — is not available, and the reason
is subtler than an absence of machinery.** The machinery exists; it is
**self-referential**, so it cannot detect what it appears to guard.

`ReductionAlgorithmVersion` (`bundle/src/ids.rs:291`) is a wire field of the
bundle superblock (bytes `68..72`, `superblock.rs:20`), and there **is** a
production writer path:

- `reduction_version_for` (`bundle.rs:989`) sets a new superblock's version from
  **the canonical base's own reported version**, or `default()` (zero) when
  there is no base — its doc: *"only a base records a reduction."*
- `open` (`bundle.rs:396`–`:399`) rejects a bundle whose base's version
  **disagrees with the superblock's**.

So a document's declared reduction version is sourced from the base and then
checked against itself. **Nothing anywhere compares either value against the
semantics the running implementation actually implements.** A base reduced under
old semantics carries its own version forward, agrees with the superblock it
seeded, and is accepted. **The check is not vacuous** — it catches a corrupt or
tampered base whose version disagrees with its superblock. What it cannot catch
is a *valid* stale base: one whose version was conformingly propagated. So the
check **necessarily passes for a conformingly propagated stale base**, which is
exactly the case `core_spec.tex:11614` exists to prevent.

*(An earlier draft of this pin claimed "every construction site is an unrelated
hardcoded literal." That was false, and the error is instructive: the search
behind it looked for `ReductionAlgorithmVersion(` constructor calls, which by
construction cannot find a path that propagates an existing value without
constructing one. The instrument could not observe the thing it was used to
rule out.)*

Two supporting facts stand: there is **no constant or accessor naming the
implementation's current reduction semantics**, and **`ids.rs:288`–`:289` states
that "the algorithm catalog itself lives in `epiphany-ops`" while nothing of the
kind exists there** — a doc comment asserting a false fact about another crate,
which is **P13-S26's pattern, second instance.**

**Scope of the claim, deliberately narrowed:** this inspection establishes that
**the current implementation has no mechanism** to detect a reduction-semantics
change. It does **not** establish that no such change in the project's history
was ever detectable; that would need a history audit this rung has not done, and
the stronger sentence must not be written into the ledger.

**Therefore this rung does NOT invent the missing machinery**, which would be a
scope explosion and a separate design — **and, because it does not, this rung
cannot execute.** Pin 0 records the break; it does not discharge it. S16 waits on
S27's disposition. What pin 0 still requires of the eventual rung:

1. **State the break** in `spec/PASS13_CANDIDATES.md`'s P13-S16 row — which
   stays **blocked on P13-S27**, not resolved — and in `operation_catalog.tex`'s
   Revision History entry: canonical bases materialized before this rung must be
   **rebuilt**, not reused, because `create_staff` now writes
   `StaffGroup.members` and `create_staff_group` now refuses inputs it
   previously accepted.
2. **File P13-S27** — that the reduction-version machinery is
   **self-referential**: `reduction_version_for` (`bundle.rs:989`) sources a new
   superblock's version from the canonical base's own self-report, and `open`
   (`bundle.rs:396`) checks only that the two agree, so **the current
   implementation has no mechanism comparing either against the semantics it
   actually implements.** Supporting: no constant or accessor names the current
   semantics, and `ids.rs:288`'s catalog claim is false. **The stronger
   historical claim — that no such change has ever been detectable — is NOT
   established by this inspection and must not be written.** This rung is the
   occasion, not the cause.
3. **Assert nothing it cannot enforce.** No test may claim stale bases are
   rejected; nothing rejects them. The break is recorded, not guarded, and the
   contract says so plainly rather than implying coverage.

**Pin 1 — refuse a non-empty carried `members`, using the existing helper.**
In `create_staff_group` (`reduce.rs:4458`), before the liveness loop, refuse a
carried non-empty `members` with `container_not_empty()`. Match the idiom and
comment style of `:4174`, `:4246`, `:4310`. **Do not introduce a new
`PreconditionFailureReason`**; `ContainerNotEmpty` stays at 10 and no schema
document moves.

The liveness loop (`:4489`–`:4502`) becomes unreachable for non-empty members
and **must be deleted, not left dead** — a precondition that cannot fire is the
kind of residue this pass keeps finding.

**No behavioural test can prove that deletion.** A retained dead loop passes M1
and every other assertion here, because refusal short-circuits before it. Pin 1
therefore requires a **structural gate**: read
`crates/epiphany-ops/src/reduce.rs`, slice the `create_staff_group` body
(brace-matched from its `fn` line to its closing brace, **production source
only** — never the test module), whitespace-normalize, and assert the slice
**contains** the empty-members refusal and **does not contain** any
`TypedObjectId::Staff` liveness check or `TargetMissing` construction.

**Pin 1a — three existing tests assert what pin 1 removes; all three must be
revised.** Each currently *passes* by asserting the outgoing rule:

| Test | Line | Currently asserts | Under pin 1 |
|---|---|---|---|
| `t6` | `reduce.rs:16154` | `CreateStaffGroup.members` naming a non-live target is `TargetMissing`, one of three referential loops asserted separately | that arm becomes `ContainerNotEmpty` and stops testing a *referential* loop at all; the `CreatePartDefinition.staves` and `CreateView.active_layers` arms are untouched and must stay |
| `t7` | `reduce.rs:16229` | those same preconditions are **not** enforced base-free | pin 1's refusal is **not** graph-gated — it is a property of the carried value — so this arm now refuses base-free too, inverting the claim for that arm only |
| `t9` | `reduce.rs:16454` | a from-empty score passes `check_invariants`, and **each skipped reducer check independently makes invariant 10 fire**, on a fixture that already attempts a dangling reference in each of the four ops | the dangling-member fixture can no longer enter the graph, so both the fixture and t9's own mutation set change |

**`t7`'s inversion is the subtle one and must be stated in its doc block:** a
graph-aware precondition asks about the *universe* and cannot run base-free; an
empty-container precondition asks only about the *carried value* and therefore
runs everywhere. Different classes — conflating them is how a later reader would
wrongly "restore" the graph gate.

Each keeps its id and gains a doc note recording what it asserted before this
rung and why the assertion inverted.

**Pin 2 — maintain the projection in `create_staff`.**
In `create_staff` (`:4330`), after the graph push at `:4386` and only when
`op.staff.group == Some(g)`, append the new staff's id to `g`'s `members` **in
`self.graph`'s `score.staff_groups`**. Under base-free reduction
(`self.graph.is_none()`) there is no graph to maintain and nothing is written —
matching how `:4385` already guards the staff push itself.

Appending is set-union and order-independent per staff id; the append must be
idempotent (never add an id already present), since convergence replays.

**Pin 3 — carried and derived are kept apart, and `staff_group_values` holds the
carried value.**
`self.staff_group_values` (`:4507`) MUST continue to store the value **as
authored** — which pin 1 guarantees is always empty-membered. Pin 2 writes the
maintained members **only** into `self.graph`. The re-carry comparator at
`:4466` is not changed and must keep comparing against the carried value.

**Pin 3a — a permanent named regression test, not only mutation M3.** Add
`t8c_recarry_compares_against_the_carried_members_not_the_derived`: in one
session, `CreateStaffGroup(g, [])` → `CreateStaff(s, group: Some(g))` →
re-carry `CreateStaffGroup(g, [])`, asserting `AlreadyApplied` **and** that the
graph's `g.members == [s]` at that moment. A mutation demonstrates the hazard
once; only a test keeps it demonstrated.

**Pin 4 — base ingest reseeds the carried value, not the derived one.**
At `:1619`, seed `staff_group_values` with the group's value **with `members`
emptied**, not `group.clone()`. This is exact rather than lossy: pin 1 makes
empty the only authorable carried value, so the reconstruction is the carried
value. The comment at `:1614`–`:1618` must be extended to say why, naming the
reload hazard of §0.4 — otherwise a later reader "fixes" it back to
`group.clone()` and reintroduces a defect that only appears after a snapshot.

**Pin 4a — a permanent named regression test for the reload path.** Add
`t8d_recarry_after_reduction_onto_a_materialized_base_stays_idempotent`: reduce
the pin-3a sequence, materialize the score, **re-reduce onto that score as a
base**, then re-carry `CreateStaffGroup(g, [])` and assert `AlreadyApplied`.
This is the only test that would exercise the base-seed path; pin 4 is
unguarded without it.

**Pin 5 — close the undo hole in the unguarded direction.**
The `Staff` undo `retain` arm (`:2967`) MUST also strip the removed staff's id
from every group's `members`, in both `self.graph` and — if any derived copy is
held — wherever else the projection lives. The `StaffGroup` arm (`:2977`) needs
no change: `:6736`'s reference guard already blocks undoing a group a live staff
names.

**Pin 6 — graph invariant 21, `StaffGroupMembershipAgreement`.**
Flags both directions: a staff whose `group` names a group whose `members` omit
it, and a group listing a staff whose own `group` is not that group. Witnesses
must name both ids and the direction.

This is an **append with a documentation footprint** — `invariants.rs:149`,
`all()`, the count guard at `:6064`, and `core_spec.tex`'s normative enumeration
(which currently ends at 20 and which **P13-S26 has just established is already
under-describing invariant 10** — do not attempt to repair invariant 10's prose
here; that is S26's rung).

**The count is not the test.** `all().len() == 21` passes with the dispatch arm
deleted — the project already knows this, which is why
`m40_check_invariants_dispatches_invariant_20` exists
(`invariants.rs:6045`–`:6066`, whose own doc says *"`all().len() == 20` passes
even with the dispatch arm deleted, so this row must instead show a score
violating ONLY invariant 20 is actually flagged by the top-level
`check_invariants` entry point"*). Pin 6 requires the **same behavioural shape**:
a score violating *only* invariant 21, flagged through `check_invariants`, with
the dispatch-arm deletion as its signing mutation. A count assertion may
accompany it and may not replace it.

**Pin 6 is coupled to pin 5 and they split only together.** M5 signs the undo
hole by requiring invariant 21 to *observe* the residue. If invariant 21 is
deferred to a later rung, then either pin 5 and M5 defer with it, or M5 must be
restated to assert the leftover member **directly** on the materialized score
rather than through `check_invariants`. Splitting pin 6 out while leaving M5 as
written would leave the undo repair unsigned.

**Pin 7 — `t8b` inverts.**
Rewrite `t8b_both_permitted_stale_forms_hold` as
`t8b_the_projection_is_maintained_and_the_spurious_form_is_refused`: same two
authoring orders, verdicts inverted. Its doc block must record that it
previously pinned the opposite, and why the change is the ratified disposition
rather than a regression. **Deleting it is forbidden** — the pairing of the two
orders is the coverage.

**Pin 8 — the four G3a undo-repair tests are re-verified, not assumed.**
`:16838`, `:16992`, `:17138`, `:17345` each construct the missing-member form,
which pin 2 now makes an agreeing form. They assert on undo effects rather than
on `members`, so they are expected to survive — **expected, not known.** Each
must be run and its verdict reported; any that changes is a finding.

**Pin 9 — no byte artifact moves.**
`spec/vectors/decode_vectors.txt`, `ops/src/vectors.rs:829`'s literal-byte
vector, and `textproj/src/vectors.rs:198`'s golden are all **unchanged**,
because refusal is at reduction and decode is untouched (§0.2). Confirm, do not
assume. No schema major or minor, no accept-set move, and **no schema/binary-format
companion version bump**. This is narrower than "no version bump": pin 10 *does*
require an Operation Catalog version bump and Revision History row, because the
catalog's normative text about these two operations changes.

**Pin 10 — specification surfaces.**
`operation_catalog.tex` §CreateStaff and §CreateStaffGroup currently state the
disposition-B stale-form semantics normatively (`:1265`–`:1278`, `:1531`–`:1540`)
and must be rewritten to the maintained rule, with a Revision History row and a
version bump. `core_spec.tex`'s `Staff.group` / `StaffGroup.members` docs
(`:5585`–`:5591`, `:4235`–`:4241`) and the two Rust doc comments
(`graph.rs:842`–`:847`, `:1643`–`:1649`) likewise.

**The two Rust doc comments are grep-asserted** by
`graph.rs:2136` and `:2160` (needles at `:2139`, `:2142`, `:2163`, `:2166`,
`:2148`, `:2172`, `:2176`, `:2180`). Those guards must be updated in step, and
the updated needles must assert the *new* rule — a guard left asserting the old
words would fail loudly, but a guard weakened to a substring both rules share
would pass silently, which is the worse outcome.

**Each new needle MUST be wording the disposition-B comments cannot satisfy.**
Retaining only "sole authority" and "non-authoritative" is insufficient: both
phrases are true under B and under A, so a guard built from them alone passes
against the text it is meant to have replaced. Require phrases that are false
under B — for example **"maintained from `Staff.group`"** and **"must agree"** —
so that reverting either comment to its B wording fails the guard. M7 signs
exactly this.

**Pin 11 — the ledger. AMENDED 2026-08-09: the mandated state changed when S27 landed.**

> **This pin required the P13-S16 row to read "blocked on P13-S27".** S27 landed and was
> accepted at `4df8e25`, so that state is now **false**, and a pin mandating a false
> ledger state would put this contract in contradiction with the ledger it governs.
> **As amended, the row must read: UNBLOCKED — S27 landed; this contract is DRAFT and
> needs ratification before dispatch.** Everything else the pin requires recorded is
> unchanged.
>
> The second paragraph's instruction to **file P13-S27 in the same edit** is
> **discharged** — S27 has its own row, its own ratified contract, and its own
> `RESOLVED — IMPLEMENTED 2026-08-09 (pin 10)` marker. It is no longer this rung's to
> file.

`spec/PASS13_CANDIDATES.md`'s P13-S16 row → **UNBLOCKED, S27 landed** *(was: blocked on
P13-S27)*, recording the
`ContainerNotEmpty` reuse (and that a new reason was considered and proved
unnecessary), the base-ingest hazard, the `t8b` inversion, the `t6`/`t7`/`t9`
revisions, and — per pin 0 — that canonical bases materialized before this rung
must be rebuilt rather than reused.

**File P13-S27 in the same `spec/PASS13_CANDIDATES.md` edit** — it is a
prerequisite discovered by this rung, not independent ledger cleanup, and the
two rows must land together so the blocking relation is visible from either end.
Its claim, at the scope §0's inspection supports: the reduction-version
machinery is **self-referential** — `reduction_version_for` (`bundle.rs:989`)
sources a new superblock's version from the canonical base's own self-report,
and `open` (`bundle.rs:396`) checks only that the two agree — so **the current
implementation has no mechanism comparing either against the semantics it
actually implements**, and `core_spec.tex:11614`'s rebuild requirement is
unenforced. Supporting: no constant or accessor names the current semantics, and
`ids.rs:288`'s claim that the catalog lives in `epiphany-ops` is false — a
second instance of **P13-S26**'s pattern.

**Do not write the stronger historical claim** ("no reduction-semantics change
has ever been detectable"); §0 does not establish it.

**The P13-S16 row does NOT move to RESOLVED in this edit.** It records the
disposition-A plan, names this contract, and is marked **blocked on P13-S27**.

---

## §2. Touch table

| # | File | Change |
|---|---|---|
| 1 | `crates/epiphany-ops/src/reduce.rs` | pins 1, 1a, 2, 3, 3a, 4, 4a, 5, 7, 8; and `create_staff_group`'s own doc comment, which states the disposition-B rule |
| 1b | `crates/epiphany-ops/src/payload.rs` | `CreateStaffGroupOp`'s doc (`:1789`ff) states that graph-aware reduction preconditions every carried member resolves to a live `Staff` and that "the mint stores `members` exactly as given and neither maintains nor trusts it" — **both clauses become false** |
| 1c | `crates/epiphany-ops/src/valuegen.rs` | `staff_group`'s doc (`:372`–`:375`). **Precision:** the helper still carries and preserves supplied `members` exactly, and "**never normalizes**" stays true and MUST be preserved — the helper is unchanged. What becomes false is the framing that such a non-empty value is "the value a `CreateStaffGroup` mints" (it can no longer mint one), and the disposition-B attribution. Rewrite those two clauses only |
| 2 | `crates/epiphany-core/src/invariants.rs` | pin 6 |
| 3 | `crates/epiphany-core/src/graph.rs` | pin 10 (two doc comments + their two grep guards) |
| 4 | `spec/operation_catalog.tex` (+ `.pdf`) | pin 10 |
| 5 | `spec/core_spec.tex` (+ `.pdf`) | pins 6, 10 |
| 6 | `spec/PASS13_CANDIDATES.md` | pin 11 |

Regenerate the two PDFs **only after** their sources reach final form.

---

## §3. Mutation plan

Applied, **run**, output recorded verbatim, restored **by hand-editing back**.

**M1 — the refusal fires.** Remove pin 1's emptiness check; the rewritten `t8b`
spurious-order assertion must fail.

**M2 — the maintenance fires.** Remove pin 2's append; the rewritten `t8b`
missing-order assertion must fail.

**M3 — the re-carry stays idempotent.** Make pin 3 write maintained members into
`staff_group_values`; a byte-identical re-carry must degrade from
`AlreadyApplied` to `RecreateContentMismatch`. **This mutation must be observed,
not reasoned about** — it is the sub-pin disposition A named.

**M4 — the base-ingest hazard is real.** Restore `:1619` to `group.clone()`;
pin 4a's test must fail. **If it does NOT fail, pin 4 is unmotivated and that is
a finding** — report it rather than keeping a guard nothing needs.

**M3 and M4 are signed by pins 3a and 4a, not by manual demonstration.** Run each
mutation against its named test so the guard, not the transcript, is what
survives the rung.

**M8 — the dead loop is actually gone.** Reinstate the member-liveness loop in
`create_staff_group` **after** pin 1's refusal, where it cannot fire. Every
behavioural assertion in this contract must still pass, and the pin-1 structural
gate must fail. This is the only signature available for a deletion that no
behaviour observes.

**M9 — `t7`'s inversion is real, not assumed.** With pin 1 in place, run the
`CreateStaffGroup` arm of `t7` base-free and confirm it refuses; then graph-gate
pin 1's refusal behind `self.graph.is_some()` and confirm that arm reverts to
applying. Signs that the empty-container precondition is deliberately
universe-independent.

**M5 — the undo hole is closed.** Remove pin 5's strip; undoing a `CreateStaff`
must leave a disagreeing pair that invariant 21 flags.

**M6 — invariant 21 sees both directions.** Delete each arm in turn; each
deletion must leave a distinct disagreeing fixture unreported.

**M7 — the doc guards discriminate.** With pin 10's needles updated, revert one
doc comment to its disposition-B wording; the guard must fail. A guard that
passes against both wordings is weakened, not updated.

---

## §4. Gate

1. `cargo test --workspace` — full pass. **The count will move** (t8b renamed,
   new invariant tests). Report the new total and the delta with its cause.
2. `cargo clippy --workspace --all-targets -- -D warnings` → clean.
3. `cargo fmt -p epiphany-ops -p epiphany-core --check` → clean.
   **`cargo fmt --all` is forbidden.**
4. `git diff --cached --check` → clean, after staging; confirm the staged list
   is exactly §2.
5. `spec/vectors/decode_vectors.txt` **unmodified** (pin 9). Confirm by `git
   status`, not by inspection.
6. Invariant 21 is reached **through `check_invariants`** on a score violating
   only it, in the shape of `m40_check_invariants_dispatches_invariant_20`
   (`invariants.rs:6045`). `all().len() == 21` and the `core_spec.tex`
   enumeration ending at 21 are checked **in addition**, never instead.
7. The four pin-8 tests each run, each verdict reported.
8. The pin-1 structural gate: `create_staff_group`'s production body contains
   the empty-members refusal and no member-liveness/`TargetMissing` path.
9. `t8c` (pin 3a) and `t8d` (pin 4a) both present and passing, by name.

---

## §5. Staging and boundary

Stage only §2's files, by explicit path. **Never `git add -A`.**

**A concurrent session commits here.** Re-check `HEAD` before staging and before
commit. **Never** `git reset`, `git restore --staged`, `git checkout`, or `git
stash`.

**Out of bounds — MUST NOT be read, written, or staged:** the entire `spikes/`
tree, `spec/PLAN_EDITOR_APP.md`, `spec/CONTRACT_EDITOR_*.md`,
`spec/ANALYSIS_GENESIS_PERSISTENCE.md`, `spec/ANALYSIS_TEXT_RUN_PRIMITIVES.md`,
`spec/DRAFT_T4_FIXTURE_RECIPE.md`, `crates/epiphany-render-svg/**`,
`crates/epiphany-glyphs/**`, `crates/epiphany-editor-gui/**`,
`crates/epiphany-testkit/benches/editor_pipeline.rs`, the root `Cargo.toml`,
`.claude/worktrees/`.

**Ratified contracts MUST NOT be edited**, including
`spec/CONTRACT_GENESIS_G3A_ENTITIES.md` and `…_G3A_UNDO_REPAIR.md`.

**Do not repair `core_spec.tex`'s invariant-10 prose** — that is P13-S26, and
its evidence at `invariants.rs:69`–`:71` must stay intact.

**The executing agent MUST NOT commit.** Leave the work staged.

---

## §6. Report requirements

1. The nine mutations (M1–M9), each with verbatim failure output.
2. The nine gate results, each with its command.
2a. For pin 0: confirmation that **nothing** was added claiming to reject or
   detect stale canonical bases, and that the break is recorded only in prose.
3. The staged file list, and the test-count delta with its cause.
4. The four pin-8 verdicts, and the `t6`/`t7`/`t9` revisions with what each
   asserted before and after.
5. Anything contradicting this contract. A contract defect reported is worth
   more than a contract satisfied.