check: literal-defaulting cleanup follow-ups (post #9605)

[jaredramirez]

Follow-up cleanup to #9605 ("check: lots of literal things"), driven by a multi-pass review (correctness, duplication, and a local-optimum/architecture audit). Sixteen well-scoped commits, each behavior-preserving except the first (a latent bug fix). No snapshots were harmed in the making of this PR

Correctness

• fix(emit): parenthesize negative typed numerals under unary operators — unaryReceiverNeedsParens only guarded e_num_from_numeral, so a unary op over a negative e_typed_num_from_numeral re-emitted as --5.U8 (re-tokenizes as binary minus). Merged the two from-numeral arms; added a round-trip unit test.

Duplication squashes (Check.zig)

• extract collectConstraintSignatureReachable — the signature-footprint walk was copy-pasted in boundaryDefaultLeaksIntoSignature and runLiteralDefaultingRounds; these must agree or a literal could be partitioned non-interfering yet warned.
• share numeral-probe digit-fit and verify loops — tryCommitNumeralCandidate and commitLiteralGroupDefault now call shared rangeNumeralDigitsFit / candidateSatisfiesRangeConstraints instead of hand-synced copies.
• centralize boundary-leak warning bracketing — new commitGatheredGroup, symmetric to commitGatheredLiteral, collapses the inlined before/commit/after warning dance.

Consolidation

• derive CheckedStaticDispatchConstraint flat fields via accessors — drop the redundant binop_negated/num_literal (re-derived from origin at 5 sites) in favor of binopNegated()/numeralInfo(), finishing the consolidation the PR began on the type-store side.

Architecture audit findings

• single source of truth for binding powers — RocEmitter kept a hand-synced parallel copy of the parser's binding-power table (its comment falsely claimed it "mirrored exactly"; the numbers were actually off-by-2). A parser precedence change could have silently made re-emitted operator expressions reparse to a different tree. The emitter now maps Binop.Op → Token.Tag and looks up the parser's one table, and a comptime assertion enforces that every Binop.Op resolves to a token with a defined binding power (so the lookup's .? can never panic — a bad future mapping becomes a compile error).
• centralize the literal-kind tie-break — numeral > quote > interpolation was replicated at three sites synced only by a doc comment; now one StaticDispatchConstraint.dominantLiteralKind kernel.
• buffer .eql edges in the generalization-boundary scan — defaultLiteralsAtGeneralizationBoundary re-scanned the whole module constraint list each fixpoint round; now collects the .eql edges into a front-loaded reusable buffer once per call.

Tidy-ups

• probe bench counters → per-instance + debug-only — moved off process-global pub vars onto Check fields, gated every increment behind comptime std.debug.runtime_safety so release builds compile them out to a true no-op.
• SurfaceOrigin encode/decode comptime guard + round-trip test; TestEnv.initReportBuilder helper (collapsing 6 sites, hoisting one out of a loop); stale doc-comment fixes.
• Lifts all inline allocations to the root struct in Check for predictable memory use

Verification

Every commit was implemented and independently reviewed (spec + code-quality) per task, then a final holistic review of the whole stack. Each behavior-preserving change asserts zero test/snapshots drift after regeneration. Full stack: 2608 Zig unit tests pass, multi-backend eval passes, zero snapshot drift. (The binding-power comptime guard is compile-time-checked and adds no runtime behavior.)

Note: local zig build minici fails only in build-coverage-tools (the vendored kcov tool can't find system libs z/curl in this environment) — unrelated to these changes and identical on main.

🤖 Generated with Claude Code

[Anton-4]

No snapshots were harmed in the making of this PR

😄

[Anton-4]

Nice and clean :)