feat: run the new grammar engine behind an experimental flag (#467 step 5)

[SJrX]

What

Wires the new list-of-successes engine into the actual InvalidValue inspection — but behind an opt-in flag, so it can be exercised in the real IDE while the original SyntacticMatch/SemanticMatch path stays the default for everyone else.

Stacked on #471 (issue-345-4). This is the first PR where parse() does something user-visible.

Why a flag

Until now the new engine lived entirely beside the old one with unit tests. A flag (mirroring the existing Podman experimental toggle) turns the scary all-or-nothing migration into "flip it, try it, iterate," and lets the real inspection tests run against the new engine end-to-end — a much stronger correctness signal than parallel unit tests.

Changes

• ExperimentalSettings — project-level PersistentStateComponent with useGrammarParseEngine, its own storage (kept out of PodmanQuadletSettings). A checkbox is added to the shared "systemd Unit Files" settings page.
• GrammarOptionValue — when the flag is on, generateProblemDescriptors runs validate() and maps ParseOutcome onto the same problem descriptors as the old path:
  • Valid → nothing,
  • SemanticError → highlight the bad token + literal-replacement quickfixes,
  • SyntaxError → highlight from the furthest offset.
    Flag off → today's code, untouched.
• EDT safety, kept IntelliJ-free in the engine — validate() gains a pure step cap (fails open to Valid rather than flag what it couldn't fully explore) and an onStep callback; GrammarOptionValue passes ProgressManager::checkCanceled, so cancellation works without any IntelliJ type leaking into the grammar package.
• Test — GrammarParseEngineInspectionTest drives the RestrictAddressFamilies= inspection end-to-end with the flag on (valid → 0 highlights, invalid → 3), plus a flag-off sanity check. It resets the flag in tearDown because the light-test project is shared across classes.

Notes

• As discussed, the whole inspection suite is not dual-run under both engines — the engine itself is covered by ParseTest; this just proves the wiring + ParseOutcome→descriptor mapping.
• Known minor difference to reconcile before the flag could ever become default: the new engine's bad-token localization can differ slightly from the old engine for some grammars (surfaced as a quickfix range diff, e.g. ::tcp vs :tcp, in a CGroupSocketBind test when the flag leaked — now fixed via tearDown). Validity detection matches; only the highlighted range/quickfix offset can differ.

Try it

./gradlew runIde, enable Settings → Tools → systemd Unit Files → "Use the new grammar engine for value validation (experimental)", and edit a .service file.

Refs #467 #345

🤖 Generated with Claude Code

[github-actions]

Test Results

1 121 tests  +3   1 121 ✅ +3   56s ⏱️ +7s
  296 suites +1       0 💤 ±0 
  296 files   +1       0 ❌ ±0 

Results for commit 26877c2. ± Comparison against base commit 44b6e41.