feat(t27b): tri t27b reduce -- ddmin fixture reducer, decisions in reduce.t27 (Closes #6445)

[gHashTag]

Closes #6445
Refs #6488 (sieve C3), #6434, #6063

What

tri t27b reduce — the fixture reducer. It shrinks a spec to a minimal
regression fixture with delta debugging (ddmin after Zeller) over top-level
declarations, so every t27b blocker family can carry a 20-line reproduction
instead of a 384-line spec.

Two modes:

• --family X — keep shrinking while t27c (the reference) passes the
  candidate and t27b's FIRST unsupported construct is still family X;
• --disagree — keep shrinking while t27c passes and t27b FAILs, MISMATCHes
  or CRASHes (a shared block is not a disagreement).

Where the decisions live

All of them in specs/tri/t27b/reduce.t27, compiled to
gen/c/tri/t27b/reduce.c (L2) and called through ctypes — the same pattern as
specs/tri/t27b/dogfood.t27 / t27b_dogfood.py (#6457):

• the interestingness table (interesting, startable);
• the ddmin state machine (running, removal, next_cur, next_gran,
  next_pos, start_gran, chunk_mask, count_bits, span);
• the budgets (MAX_UNITS = 64, MAX_CANDIDATES = 512, budget_done);
• the recorded result of a real reduction (constants + test, below).

scripts/tri_loop/t27b.py (owner-approved exception, unmodified policy) keeps
only I/O: unit splitting, running t27c/t27b, a content-hash oracle cache, and
writing the fixture. The cache key carries the oracle binaries' hashes.

The reference predicate (a trap this PR closed)

The first run produced a 21-line fixture that t27c typecheck accepted — but
which referenced undeclared identifiers (GAT_POS), so t27c test-report
said BLOCKED does not compile while still exiting 0 (#6788 made test-report
print BLOCKED instead of a vacuous all-pass). The oracle now requires:
typecheck exit 0 AND test-report exit 0 with no BLOCKED and no FAIL count.
Typecheck alone is not a reference pass.

The recorded run (test recorded_gfternary_reduction)

specs/numeric/gfternary.t27 reduced with --family "type (alias)":
45 units / 384 lines → 5 units / 30 lines in 428 candidates (84 of the 512 budget left),
inside every budget; the fixture's first t27b blocker is still type (alias).
The test asserts the measured numbers, so the done-when cannot drift.

Verified locally: spec 5/5 tests, t27c gen-c regenerated and deterministic,
tri t27b reduce byte-identical on a rerun, negative controls honest
(not-interesting original → exit 3; --disagree on a blocked spec → exit 3;
--budget N stops and says so; >64 units → exit 2), and the five
scripts/ci/test_the_t27b_*.py / test_a_t27b_*.py suites pass.

🤖 Generated with Claude Code

[gHashTag]

Closing as a duplicate of #6830, which landed while this implementation was in flight (the cold reduction run alone took ~45 min; #6830 merged 13:26Z, this PR opened ~14:00Z).

The landed version is the better keeper: it reduces at three levels (declarations, statements, expressions) where this one stops at declarations, and its reference guard compiles candidates through t27c's Zig backend, which rejects undeclared identifiers outright — the exact trap this PR's oracle hardened against by parsing test-report output for BLOCKED (bare t27c typecheck accepts undeclared identifiers, and #6788 made test-report print BLOCKED while still exiting 0). Two implementations would give one command two homes; keeping theirs.

For the record, this run reached the same endpoint on the same target: specs/numeric/gfternary.t27 --family "type (alias)" shrank 45 units / 384 lines to 5 units / 30 lines in 428 candidates — the family fires on the alias use in a signature, not on the bare alias.