docs(findings): verify the choreography fixtures; record the --emit-core-ast verdict drift

[Ch4s3]

What this is

Docs only. No ledger change, no checker change.

I set out to mirror march's new choreography/endpoints fixtures into
scripts/expected-skips.txt. They cannot be added yet, for a reason this repo
already documented and PR #24 already hit: conformance.yml pins march
(binary and corpus) at 6867c783 (2026-08-08, now 1057 commits back), and
none of these files exist in that checkout, so a ledger entry for any of them
is a stale-entry SKIP-LEDGER MISMATCH against CI's actual corpus.

So instead of a ledger edit that would turn CI red, this records the
verification and the measured cost of the pin bump that would let it land.

Fixtures verified (march main HEAD 3ebe6c17, march-lean-check at this repo's HEAD)

Every one SKIPs cleanly. None is a MISMATCH; none needs a
known-limitations.txt entry.

fixture	why it skips
accept/t270_endpoints_labelled_steps	out-of-fragment construct in a declaration
accept/t273_crash_branches_logging	out-of-fragment construct in a declaration
reject/t271_endpoints_label_on_branch_head	out-of-fragment construct in a declaration
reject/t272_endpoints_label_msg_prefix	out-of-fragment construct in a declaration
reject/t274_crash_receive_without_branch	out-of-fragment construct in a declaration
reject/t275_crash_branch_on_reliable_sender	out-of-fragment construct in a declaration
reject/t276_crashed_role_in_own_crash_branch	out-of-fragment construct in a declaration
reject/t277_crash_third_party_not_told	out-of-fragment construct in a declaration
reject/t278_crash_choose_two_detectors	out-of-fragment construct in a declaration
reject/t279_may_crash_unknown_role	out-of-fragment construct in a declaration
reject/t263_endpoints_payload_without_json_codec	out-of-fragment construct in a declaration
grammar/parse/p38_protocol_labelled_message_step	out-of-fragment construct in a declaration
grammar/reject/r17_protocol_label_after_arrow	no module (parse failure)

There is no accept/t263_… payload-codec fixture — accept-side t263 is
t263_linear_opt_in_second_param, unrelated. The payload-codec fixture is the
reject-side one above.

The ledger lines are recorded verbatim in specs/march-findings.md, padded to
the file's column, ready to paste when the pin bumps.

New march-is-wrong finding: --emit-core-ast's verdict disagrees with --check

bin/main.ml binds has_user_errors from the diagnostic set and hands it to
Emit_core_ast.run as ~rejected: before the --check path lowers to TIR
to judge cap no_alloc allocation contracts and the stdlib-mediated
capability-ceiling checks. Three reject fixtures therefore emit
"verdict": "accept" while march --check on the same file errors out:

$ march --check specs/lang/types/reject/t43_cap_no_alloc_tuple.march
-- ERROR -- `make_pair` is in `cap no_alloc` module `NoAllocPair` but allocates.
$ march --emit-core-ast specs/lang/types/reject/t43_cap_no_alloc_tuple.march | jq -r .verdict
accept

Also reject/t180_ceiling_stdlib_mediated_under_check and
reject/t182_ceiling_module_let_stdlib_mediated. The hoist comment at that
binding claims it is "the same accept/reject condition --check uses below";
for these three it is not. The harness's step-4 self-consistency cross-check
is precisely what caught it — it reports MARCH_SELF_INCONSISTENT, a hard
failure with no ledger to absorb it. Not yet reported upstream.

Measured cost of a pin bump to HEAD

Full run, --corpus-dir + --lang-dir at march main 3ebe6c17:

total files 498 | MATCH 97 | MISMATCH 1 | ERROR 0 | SKIP 396 (accept 189, reject 207)
KNOWN_LIMITATION 3 | MARCH_SELF_INCONSISTENT 3 | CORPUS_VIOLATION 0
SKIP-LEDGER delta: +125 newly skipping, 1 stale

The single stale entry is accept/t77_refine_hof_bypass_limitation.march,
which march renamed to accept/t77_refine_hof_pass_site_rejected.march.

Two blockers beyond the ledger:

1. reject/t262_toplevel_let_annotation_mismatch — hard MISMATCH. march now
   checks a module-level let's annotation against its RHS
   (let x : Int = "hello"); this checker's inference does not. The
   annotation is present in the Core AST, so it is not a
   known-limitations.txt candidate (that file is gated to rejections erased
   from the Core AST) — it is a real checker gap to fix.
2. The three MARCH_SELF_INCONSISTENT files above. No ledger category exists
   for them and the cause is march-side, so nothing in this repo clears them.

A green bump therefore needs a checker fix for (1) and an upstream fix for
(2). Both are out of scope here.

CI

Docs-only diff, so the gate runs against the unchanged pin and stays green.