`docs/10-effects.md` said the compiler checks the row law. It did not. E0506
asked whether a row entry DENOTES and nothing asked whether the claim was
TRUE, so a body could incur `Log` with no row, or name `Log` and never log,
and both compiled. Both halves are now refused:

    direct ∖ declared  ->  E0413   incurred and not named
    declared ∖ direct  ->  E0509   named and not incurred

THE LAW LIVES WITH THE GRAPH. `ck_effects::rows` is the consumer ruling #30
reserved that crate for, and it uses the `ck_diag` edge the crate had declared
in its manifest and used in zero source files since it landed. Both directions
compare against `direct_effects`, never `reaches`: comparing against the
transitive set would demand exactly the restatement the law forbids.

The transitive half is not idle — it is what makes E0509 actionable. Every real
over-declaration in this tree is a row restating a callee's effect, so the
message names the callee (`carries_only`'s question, asked through `callees`).
Where no carrier is among the fed frames it says nothing about callees at all:
the closure is app-only, so "nothing it calls incurs one" is a claim this
compiler cannot always support.

THE SEAMS ARE EXISTING ONES, NOT NEW PARALLEL WALKS.

  · `ck_syntax::effect_row_of` — one reader for a row's syntactic shape. Three
    callers were hand-rolling `child_node_of(node, EffectClause)`; the four row
    positions now share one definition and a fifth would inherit it.
  · `ck_resolve::effect_rows` — collects `FnId -> WrittenRow` off `module_fns`,
    already the canonical "every function this module declares". Actor slots are
    covered without being named (the §2a desugar hands them over as impl
    methods) and bodiless declarations are excluded without an exemption
    (`module_fns` skips traits and effect ops, so the law inherits the right
    domain instead of allow-listing to it). A fn with no row maps to an EMPTY
    row, never an absent key: "I incur nothing" is a claim, and it is the one
    E0413 most often refutes.
  · `ck_driver::effect_rows::Frames` — feeds the fixpoint from
    `ResolvedBody::resolutions`, the same map `interface_ops` reads, so provider
    DI and the effect graph cannot disagree about what a call is.

An under-declaration anchors on the declaration's own name token, not on
`fn_data(id).span`: N1 mints the `FnId` from the `.ckh`, so the interner's span
points at the header while the body is in the `.ck`. The row is part of the
SIGNATURE, though — E0403 checks impl against header — so a fn whose body incurs
an effect declares it in both files or in neither.

THE MIGRATION: 301 rows added, 9 removed, 127 files. Every removal was one
shape, a row restating an effect a callee incurs, and
`tests/corpus_test_run/greeter` is where that shape had been planted
deliberately all along — the cell cited it at its pre-`0d6d1c13` path, which no
checkout contains.

TWO DEFECTS THIS SURFACED, NEITHER OF THEM THE ROW LAW'S:

  · `grammar_dsl/measure.rs` scanned `delimited_after(src, "effect", '{', '}')`,
    which matches the `effect` of a ROW and then measures the function's body as
    a declaration body. Corpus-wide rows made it report M14 = 24 at a `test fn`.
    Guarded on declaration position; `Grammar.ckx`'s table needed no change,
    because the measurement was false rather than shifted.
  · Spans on a desugared page point into generated text — an existing E0506 on a
    row at real line 28 reports 108, the same shift. Pre-existing and left
    alone, but E0413 will be the most-read diagnostic on those pages, so it
    wants its own slice.

`effect_row_is_not_checked_yet` is renamed to
`effect_row_is_checked_in_both_directions` (a rename, so `CELL_COUNT` holds at
440) and declares `typecheck: E0413, E0509, E0509` — the three violations its
page carries, and not redundant: `quiet` falls to the direct set alone, while
`spawn` really is on a path to `Log` and falls only to the rule that a row never
restates a callee's.

The cell carried no `// STATUS:` line while its own prose said it was BROKEN, so
it sat in the unmarked set — invisible to `grep -l '^// STATUS: BROKEN'`, which
is the undercount `UNMARKED_CELL_COUNT` warns about. Marking it is what put it
in the fix queue, and it came straight back out: `BROKEN_CELL_COUNT` 44 -> 45 ->
44, `UNMARKED_CELL_COUNT` 146 -> 145.

`ck_driver/src/load.rs` hit the R3 cap, so the effect-graph seam split to its own
module. `ck_driver/src/lib.rs` went 1322 -> 1323 for one `mod` line; that red is
still the standing one and still wants the driver-root split, not a cleanup.

Gate: 4 legs green. Workspace sweep 1610 passed, 1 failed — the standing R3.
`clippy --workspace --all-targets -D warnings` 0, `fmt` clean.

@@ -0,0 +55,4 @@
+/// E0506 was checked over, which is what keeps the two passes agreeing about
+/// what an entry means.
+#[must_use]
+pub fn collect_declared_rows(