feat: frontier/expected-set for error localization (#467 step 3)

[SJrX]

What

Adds a frontier: a high-water-mark recorder threaded through parse() so that even when no path matches, we know where parsing got stuck and what was expected there.

Stacked on #469 (issue-345-2). This PR's diff is the Frontier type, the one extra (defaulted) parse() parameter threaded through the combinators, and the updated validate()/tests.

The problem it fixes

parse() only returns successful matches, so a malformed value's failing paths vanish and we lose how far we got. For AF_INET, AF_INET6 against the real grammar Seq( …family-list… , EOF() ):

1. the family-list consumes AF_INET and stops at the comma (offset 7);
2. the outer Seq tries EOF() at 7, which matches nothing → the whole Seq yields an empty sequence;
3. validate()'s maxOfOrNull { it.end } over an empty sequence → furthest = 0.

The "we reached offset 7" fact existed for a moment and was thrown away. (This is the limitation that was pinned with a furthest == 0 assertion in #469.)

How the frontier works

A small mutable Frontier rides along through parse() (one extra parameter, defaulted so the 2-arg call still works). Every leaf matcher — the terminals and EOF — reports itself when consulted at an offset. The frontier keeps only the deepest offset reached and the set of matchers wanted there. Combinators just thread the same instance into their children; they don't record.

validate() reads it: SyntaxError now carries (furthest, expected).

One mechanism, two payoffs

• Error localization: AF_INET, AF_INET6 now reports furthest = 7 (the comma) and expected = {whitespace, end-of-input}. The pinned == 0 assertion from feat: add parse() (list-of-successes) to the grammar combinators (#467 step 2) #469 is flipped to == 7.
• Seed of completion (Grammar Based Completion #343): the expected set at a position is "what could come next here." A new test shows the empty value reports expected containing the address-family names (FlexibleLiteralChoiceTerminal) and none/~ (LiteralChoiceTerminal) — i.e. the completion list at offset 0.

Scope / notes

• The old SyntacticMatch/SemanticMatch engine and GrammarOptionValue are untouched; full suite is green.
• The expected set is typed Set<Combinator> (so EOF can appear alongside terminals); a real completion contributor would map these to lookup strings (e.g. a LiteralChoiceTerminal's choices) — that's the next step, not this one.
• Still not wired into IntelliJ; the path-cap + cancellation hook remain prerequisites before that.

Refs #467 #345 #343

🤖 Generated with Claude Code

[github-actions]

Test Results

1 118 tests  +1   1 118 ✅ +1   49s ⏱️ ±0s
  295 suites ±0       0 💤 ±0 
  295 files   ±0       0 ❌ ±0 

Results for commit 6a8c012. ± Comparison against base commit c3b8f6d.