docs(roadmap): rework v0.6.x follow-ups and demote lit-aware features

docs/ROADMAP.md

@@ -8,10 +8,12 @@ For shipped milestones and tag-by-tag release history, see [CHANGELOG.md](../CHA
 
 ### v0.6.x follow-ups
 
-- **Investigate lit-aware highlighting for `// RUN:` lines.** MLIR / PDLL / TableGen tests commonly use LLVM lit `RUN:` directives, but those directives are not plain bash: they support lit substitutions (`%s`, `%t`, `%{...}`, `%(line)`), continuations, and test-suite-specific tool substitution. Explore whether `highlights.scm` / `injections.scm` can highlight command-like structure without misrepresenting lit syntax as generic shell. Prefer explicit lit substitution captures and only inject `shellscript` into command spans if it behaves correctly on real LLVM / MLIR tests.
-- **Investigate lit-aware runnables for tests.** Zed `runnables.scm` can surface runnable syntax, but executing a single `RUN:` line correctly requires lit substitution and test-suite configuration, not just passing the comment payload to a shell. Start with a runnable for the whole current lit test file (for example, `llvm-lit $ZED_FILE`) and evaluate whether a small lit-aware helper is needed before exposing individual `RUN:` pipelines.
+- **Highlight & query polish across the three languages.** Work the concrete queue from the 0.6.x query review: PDLL user-constraint references in type position, declaration/tuple parameter names, operation-attribute keys, and call-target capture consistency; TableGen string/code delimiter policy and optional `include` outline entries; MLIR outline coverage for `#`/`!` alias definitions and external resources (pending grammar-side `name:` fields for aliases). Further edge cases join this queue as they surface.
+- **TableGen `[{ ... }]` auto-close and matching.** The merged `[{` / `}]` code-block tokens are not declared as a bracket pair. Test a config-only entry first (`{ start = "[{", end = "}]", close = true, newline = true }`); only pursue a grammar change that exposes queryable open/close nodes if Zed cannot match a multi-character pair.
+- **Query smoke tests against the pinned grammars.** Suite CI checks the Rust extension but never compiles `languages/*/*.scm` against the exact grammar SHAs in `extension.toml`, so a grammar bump can leave a stale node name unnoticed. Add a gate that compiles every query and parses representative fixtures for all three languages.
+- **Verify include navigation end-to-end.** PDLL and TableGen servers return `documentLink` for includes, while go-to-definition on the include itself may legitimately return nothing; TableGen cross-file definition jumps also had rough edges in earlier manual testing. Re-test with a correct compilation database / `extra_dirs`, capture the protocol traffic, and isolate whether a failure comes from symbol lookup, the returned URI/range, or Zed's rendering before changing anything.
+- **Document a lit workflow instead of building lit-aware features.** Running the current lit test from Zed needs only a user-defined task such as `llvm-lit $ZED_FILE`; add a README recipe rather than extension code. Per-`RUN:`-line highlighting and runnables stay out of near-term scope and are tracked under [Ideas](#ideas-unscored).
 - ~~**Verify MLIR-in-C++ raw strings.** Zed's bundled C++ grammar already injects `raw_string_content` using the delimiter as `@injection.language`. Since this extension registers `mlir`, the injection should "just work" inside `R"mlir(…)mlir"`, but it has not been confirmed end-to-end. Verify with a real `.cpp` file and document the result; if the wiring is missing, file an upstream issue.~~
-- **Continued highlight & query polish.** Steady, focused improvements across the three grammars and capability files as edge cases surface.
 
 ## Mid-term
 
@@ -31,6 +33,7 @@ For shipped milestones and tag-by-tag release history, see [CHANGELOG.md](../CHA
 
 ## Ideas (unscored)
 
+- Lit-aware `// RUN:` highlighting and per-line runnables. Demoted from v0.6.x: comments are single tokens in all three grammars, so faithfully highlighting lit substitutions (`%s`, `%t`, `%{...}`) would require a dedicated tree-sitter lit grammar injected into comment content — injecting generic `shellscript` misrepresents lit syntax — and executing a single `RUN:` line requires lit substitution plus test-suite configuration. A documented whole-file `llvm-lit $ZED_FILE` task covers the practical workflow in the meantime; revisit only if a maintained lit grammar appears or user demand justifies owning one.
 - Code-lens-style links from a PDLL `Pattern` to the TableGen op definition it rewrites, pending a suitable Zed extension API surface. First verify what `mlir-pdll-lsp-server` already provides via go-to-definition.
 - Quick-fix for "missing `include`" in TableGen (auto-insert the canonical header path), most likely as an upstream `tblgen-lsp-server` code action rather than a Zed-only feature.
 - Dialect-aware highlighting inside MLIR string attributes that embed recognized DSLs. Keep this opt-in / whitelist-driven so ordinary MLIR string attributes are not over-highlighted.