feat(gqt): add a server target and run cases against omnigraph-server

[azimafroozeh]

What & why

This PR lets the GQT runner execute a case against a running omnigraph-server (omnigraph-gqt <case> --server <URL> --graph <ID> [--token <T>]) and makes the corpus the server's conformance suite. It is the runner half of the server execution surface RFC 0039 lists beside the embedded one; timing, records and identity stay with the bench. No engine, server, API or CLI code changes; no new case syntax.

• Served executor in omnigraph-gqt-core: each step travels to the route that already serves it (--- query, branch list and show to POST /query; --- mutate and the branch statements to POST /mutate; the seed and --- load to POST /load) and is judged from the answer with the in-process row and error checks. Query and mutate bodies go verbatim, so the served path hits the parser doors a user hits.
• Admission before any request, by name: --- restart, --- concurrent, seams, settings, --- expect plan, --- expect same as v1, the # traversal: pin and index-needing queries are refused. --- expect shape is observed, not judged: it holds the executor's Arrow schema against the compiler's, which no wire answer carries, and 269 of 275 cases carry one. Every request is bounded at the case budget.
• A case opts in by declaring target: omnigraph-server; without --server that environment is planned but unselected (report scope partial), so the corpus runs in-process as before. 80 of 276 cases opt in; the rest use a refused section. A declared omnigraph-server-dst is still refused, nothing implements it.
• Differential test: gqt_served_conformance in omnigraph-server boots an in-process server per opted-in case and requires a green in-process verdict and the same served verdict; gqt_served_count refuses an empty served set; a fourth GQ Logic Tests mode, served, runs both. 80 agree, 196 skipped, 0 differ.
• The test caught two executor bugs before landing (a branch name that is not a lowercase identifier must travel as a string literal; a bare show statement needs its ; on the wire, pinned by the new show_statement_served_with_its_terminator case) and one server limit the in-process path has no twin for (Notes).

Backing issue / RFC

• Relates to the accepted RFC 0039 (docs/rfcs/0039-end-to-end-benchmark.md): its environment axis names the server execution surface; this PR implements the runner side of it through the existing routes, so no API and no RFC update is owed.

Checklist

• Focused: one logical change, the GQT server target (executor, admission, flags, conformance test with its CI mode, corpus opt-in)
• Tests: gqt_served_conformance (80 served, 196 skipped, 0 differing) and gqt_served_count; runner_dispatch::server_target_admission_refuses_before_workers pins every flag combination refused before a worker spawns (--server without --graph, with --store, with bless, with --target naming another environment, a declared omnigraph-server-dst) plus bless on an engine-and-server case; the runner_config unit test pins the five served admission arms. Rust because no .gqt case can observe a flag, a spawned worker or a request.
• Public docs: crates/omnigraph-gqt/README.md (served section) and docs/dev/ci.md (GQT (served)); no docs/user/ surface changed
• Release note: changelog.d/gqt-server-target.added.md
• Invariants: reviewed against docs/dev/invariants.md, no Hard Invariant weakened, no deny-list item hit; nearest one touched is none (no engine or server code path changes; the server crate gains a dev-dependency and two tests)

Local verification

• cargo test -p omnigraph-server --test gqt_served_conformance --test gqt_served_count -- --nocapture: 276 cases, 80 served and agreeing, 196 skipped, 0 differing, 2.0 s wall (local FS, engine v2, in-process server); the count test prints the same totals
• cargo test -p omnigraph-gqt-core -p omnigraph-gqt --locked: corpus 276, runner dispatch 26, gqt-core 193, gqt 47; clippy -D warnings on omnigraph-gqt, omnigraph-gqt-core, omnigraph-server all targets; fmt
• omnigraph-gqt <case> --server <URL> --graph <ID> against a booted omnigraph-server binary: not run; the conformance test calls the served executor directly, so the CLI flag path is pinned by its refusal rows only, and a green-verdict smoke step on the server binary is the follow-up

Notes for reviewers

• POST /query and POST /mutate refuse bodies over DEFAULT_REQUEST_BODY_LIMIT_BYTES (1 MiB), a limit the in-process path has no counterpart for: issue_856_repeat_fast_forward_past_validation_budget fails served on its large mutation and keeps its in-process environments only; no limit loosened.
• Bless counts the declared in-process environments (exactly one, direct engine) instead of every declared environment, so a case declaring engine and server still blesses; a served run never blesses.
• The served CI cell builds omnigraph-server plus the gqt dev-dependencies, a feature graph the other three modes do not share, under its own cache key and 60-minute budget; its cold wall time is unmeasured until the first hosted run.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-09T16:20:57.494401Z	1defe12	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1defe12c43

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".