v1 resolve cache: always-on via temp_dir default (drop env-var gate)

[gunbai-bot]

What

GUNBC_RESOLVED_GRAPH_CACHE_DIR was introduced in #4878 but never wired into CI, so the resolved-graph cache was silently off everywhere. This drops the env-var gate: the cache now defaults to $TMPDIR/gunbc-rg-cache, with GUNBC_RESOLVED_GRAPH_CACHE_DIR still accepted as an override for explicit placement.

Why

Without the cache, resolve_entry_with_parse_cache re-runs the full v1 type inference pipeline for every unique *_test.dag entry in the discovery corpus — even though 95%+ of each closure (std/, extdeps/) is shared across entries. Profiling on the post-#5757 baseline showed 447,734ms serial resolve total across 1011 witnesses (~443ms/entry). With the cache on, each unique closure resolves once and is reused.

Correctness

Cache key = content-hash(closure sources) + content-hash(compiler binary). Invalidates correctly on any source or compiler change. Existing falsifiers from #4878 cover cold-oracle equivalence, poisoned-hit rejection, WriteOnce concurrency, and key-mismatch miss.

Impact

Eliminates most of the 447s resolve serial sum in the CI floor corpus run. Secondary effect: per-shard peak RSS drops (no large type-checked AST graphs held per shard), which will unlock a higher corpus evaluation width via the memory-aware spawn-width formula on next re-measurement.

[gunbai-bot]

Addressing the two findings:

TypeScript files (scope creep): The TS files the reviewer saw are no longer in the diff — they were restorations, not additions. The session branch had accidentally dropped them (stale base), and the restoration commit pushed after the prior review brought them back to match main exactly. The current diff shows only the three Rust files + plan/ROADMAP session overhead.

temp_dir trust model: Correct observation. Adding a follow-up code comment is the right move rather than a code change now. The content-addressed key is the load-bearing purity guarantee; the world-writable path is a permission-error risk at worst (already fail-soft via let _ =). Will add a one-liner in the next pass.

Plan doc bundling: Acknowledged — the resource-aware-scheduler.md and ROADMAP edits are session-branch overhead that will squash-merge with the cache change. Noted for future PRs.

— sent from calm-carp-204

[gunbai-bot]

On the temp-dir uid-collision finding: confirmed the risk is real on multi-user hosts. In our CI context it is not a factor — runners are ephemeral single-user containers, so no concurrent users can race on the dir. The GUNBC_RESOLVED_GRAPH_CACHE_DIR env var override is the escape hatch for any shared-host setup. I'll track XDG-based namespacing ($XDG_CACHE_HOME/gunbc/rg-cache → $HOME/.cache/gunbc/rg-cache → temp fallback) as a follow-up once this lands.

On the bundled TS/ROADMAP files: those were stale-base restorations from a prior merge divergence (the session branch predated those files landing on main). The merge-main commit pushed just now cleans the diff to the 4 intended files: resolved_graph_cache.rs, cli_run.rs, ROADMAP.md, and the plan doc.

— sent from calm-carp-204

[gunbai-bot]

Addressed finding #1 (uid collision): cache dir now suffixed with USER/USERNAME env var → gunbc-rg-cache-${user} per user, so shared-host /tmp installs don't collide on EACCES.

On finding #2 (bundled smoke test fix): not a drive-by. The prior CI run on this PR (28145032701) failed rust_tests at html_markup_smoke_runs_green because PR #5787 updated html_markup_smoke.dag to import extdeps.languages.html but didn't update the Rust test runner's source root list. The breakage was latent on main because main's CI gate short-circuits when no Rust files changed (fast path, doesn't run the full nextest suite). Our branch's Rust file changes triggered the full gate and exposed it. The fix had to ship on this branch or CI would stay red.

— sent from calm-carp-204