docs: correct CLAUDE.md statements the tree contradicts

[OmarAlJarrah]

Summary

CLAUDE.md is binding on every contributor and is read before the code, so a false statement in it
is worse than a stale comment: it is followed. Eight of them had accumulated, all in the same shape
— a hand-maintained enumeration or count that the tree moved out from under, which is the failure
mode the file's own "Derive counts; never maintain them" rule exists to prevent.

Where a count could be replaced by the command that derives it, it was. Where a fact had to stay
written down, it now names the site it can be re-derived from.

1. The JSON codec rule was wrong in both of its clauses (closesdocs: CLAUDE.md says BigVal has a custom JSON codec; it has none #277). It said "custom
   MarshalJSON/UnmarshalJSON for special forms (the IR's sum types and BigVal do this)".
   BigVal has no codec at all — it is type BigVal string with one method, String(), and needs
   none, because being a string type is what preserves precision. And the sum is asymmetric: all
   eleven TypeDef kinds have MarshalJSON, none has UnmarshalJSON; the module's only
   UnmarshalJSON is (*TypeRegistry)'s, which peeks the adjacent kind tag and dispatches.
   Dropping BigVal alone would have left the second falsehood standing, so the sentence now
   describes the asymmetry and gives the grep that lists the codecs.
2. The switch-completeness test was called "generated". Nothing in this module is generated —
   there is no go:generate directive anywhere. The kind lists are hand-written, and
   TestTypeDef_HandWrittenKindListsAreComplete is what holds them to the kinds the ir sources
   declare. Calling it generated invites a contributor to look for a generator to re-run instead of
   editing the list.
3. "openapi has thirteen" internal packages. There are fourteen (overlay/ is the one the
   count predates). Deleted rather than corrected: the paragraph below the diagram already says the
   diagram deliberately does not name them and gives the command that lists them, so the count was
   the one part of that sentence able to rot.
4. The oracle list named five oracles; harness.Check applies six. Order-invariance was
   missing. That omission was doing active harm two paragraphs later, where the two-order-diff rule
   read as though no oracle implemented it — so contributors were being told to hand-roll, per new
   construct, a diff that already runs across the corpus under testdata/. Both are corrected, and
   the reworded paragraph says when a targeted case is still worth writing: the oracle proves
   order-independence only for constructs its inputs contain.
5. cmd/morphic/ was documented as importing engine; internal/archtest's rules map — the
   stated source of truth — says {ir, engine}.
6. cmd/morphic-harness/ was absent from the layout block entirely, which was the one part of
   docs: CLAUDE.md still describes the pre-implementation repository #63's acceptance ("every package in the tree appears in the documented layout") still unmet.
7. The normative-docs list named four of eight documents.docs/emitter-design.md was missing
   despite declaring itself normative for the emitter half, as were reference-learnings.md and the
   two micro-compiler-*.md records. All are now listed, with the two micro-compiler documents
   marked as records of landed work so they are not mistaken for a proposal or a backlog.
8. Invariant 1 read "a emitter's only input" — article left behind by the generator → emitter
   rename.

Items 1 and 7 are #277 and the last open bullet of #63. Items 2-6 are not in either issue text;
they surfaced while checking the sections around them and are the same defect, so they are fixed
here rather than left to be re-derived by the next reader.

Deliberately out of scope, so it is not lost:docs/architecture.md §3 lists the same thirteen
internal packages and also omits overlay/, and README/architecture carry the cmd/morphic import
row — those belong to #64 and are being fixed there. Two further sites found while verifying are
filed rather than fixed here: #299 (docs/micro-compiler-design.md repeats the thirteen) and #300
(the "a emitter" rename residue, 22 occurrences across five files in docs/).

Test plan

No test guards prose, so every corrected claim was re-derived from the tree by command rather than
by reading, and the whole set was then run against the pre-change file to confirm each correction
was load-bearing — 11 assertions fail against origin/main's CLAUDE.md, 0 against this branch.

Claim	Command	Result
BigVal has no codec	grep -n 'type BigVal|func (v BigVal)' ir/bigval.go	type BigVal string, String() — nothing else
the sum marshals but does not unmarshal	grep -rnE 'func .*(Unm|M)arshalJSON' ir/	11 MarshalJSON (one per typeDef() implementer), 1 UnmarshalJSON, on TypeRegistry
nothing is generated	grep -rn 'go:generate' --include='*.go' .	no output, exit 1
the test is hand-written	grep -rn 'func TestTypeDef.*Complete' ir/	TestTypeDef_KindDispatchIsComplete, TestTypeDef_HandWrittenKindListsAreComplete
fourteen internal packages	ls -d compilers/openapi/internal/*/ | wc -l	14; the same 14 as keys in archtest's rules
six oracles	internal/harness/harness.go:50-88	ends at orderInvariant → OutcomeOrderDependent
cmd/morphic imports	grep -n '"cmd/morphic"' internal/archtest/arch_test.go	{module + "/ir", module + "/engine"}
every package is documented	git ls-files '*/*.go' | xargs -n1 dirname | sort -u	all 27 covered, compilers/openapi/internal/* as a class by design
every doc is listed	ls docs/*.md	8 files, 8 named

Each command the file now instructs a reader to run was executed as written from the repo root,
including go run ./cmd/morphic-harness <file> and <dir> — both report ok.

Full gate (gofmt, go vet, golangci-lint, go build, scripts/check-coverage.sh) passes; the
change is documentation only, so no goldens move and coverage is unaffected.

Closes#63
Closes#277

Closes#339