phase 0 metrics

ROADMAP.md

@@ -107,7 +107,7 @@ A flaky or green-but-broken floor means no gate protects anything — so CI is u
 - [x] **Section 1 spawn-width foundation — std.measure expressibility** (#5470/#5478) — the FLOOR family (measure_scale_fraction_floor + measure_fit_count_floor) and the demand-CEIL family completing the ceil-not-equal-floor money-pair landed, dissolving the measure unwrap→raw-arith→rewrap §3 fork. Authority: the expressibility-frontier spec #5467 [plan](docs/plans/expressibility-frontier.md)
 - [ ] **scheduling: placement + materialization inert** — `Placement`/`Materialization` modeled + witness-passing, **no live consumer** (same band as the host-ops gap, substrate side)
 - [ ] **scheduling: resource-aware (larger allocation → larger throughput)** — `spawn_width` is a side-channel derived from a static data row (`gunbc_ci_floor_measured_peak`, coarse whole_run÷concurrency estimate); goal: width derived from per-`Runnable` cost in the plan itself, self-calibrating from measured per-shard RSS. Four nodes: (A) measurement plumbing — claim_batch emits `[measurement] per-shard-peak-rss`, claim_executor collects max *(dispatched adhoc-97532cd3-dfe)*; (B) `CostEstimate.space` on `Runnable` in std; (C) scheduler reads cost → derives width, side-channel deleted; (D) calibration loop retires static data rows. Construction invariant: host memory increase → higher width → more throughput, no `.dag` edits required. [plan](docs/plans/resource-aware-scheduler.md)
-- [ ] **caching forked** — sccache live · resolve-cache **always-on (#5789 — default `$TMPDIR/gunbc-rg-cache`, CI pending merge)** · ParseTable memo live · RecordedFixture · BuildBuddy opt-in → converge on `realize(subject)` (§2 P2)
+- [ ] **caching forked** — sccache live · resolve-cache **opt-in (`GUNBC_RESOLVED_GRAPH_CACHE_DIR`; #5789 always-on default reverted — its whole-file JSON read/write OOMs the floor, must stream before it can default on)** · ParseTable memo live · RecordedFixture · BuildBuddy opt-in → converge on `realize(subject)` (§2 P2)
 
 **Host-operation band — off-fabric, unmodeled, unenforced (G1–G3, NOW):**
 
@@ -157,7 +157,7 @@ Gate: uncached non-redundant work is an ERROR, not "slow". The cache-key-from-in
   - [x] P1 honest keys by construction — warm==cold purity oracle **LANDED (#5429)**
     - [ ] P2 one door: `realize(subject)` sole API — kernel inhabits `cache_interface.dag` (#5446); ParseTable dissolution is downstream of the dsl→v2 de-fork (§5)
       - hermetic fixtures feed P2: [x] M4.1 universal hermetic corpus governance (#5236, [plan](docs/plans/m4-universal-hermetic-corpus.md)); [ ] M5 fixture-store onto one Realization kernel ([plan](docs/plans/m5-fixture-store-consolidation.md))
-      - [x] P3 **resolve-cache enable** — purity proven (616/616); #5789 makes cache **always-on** (default `$TMPDIR/gunbc-rg-cache`, no env var required; CI pending merge 2026-06-25) ← **core ask landed**
+      - [ ] P3 **resolve-cache enable** — purity proven (616/616); cache is correct but **opt-in** (`GUNBC_RESOLVED_GRAPH_CACHE_DIR`): #5789's always-on `$TMPDIR` default was reverted because the whole-file JSON read (hit) and write (miss) buffer ~11x the packed graph in memory and OOM the concurrent-shard floor. Re-enabling by default is gated on a streaming IO realization (binary + `deserialize_from`/`serialize_into`, no whole-file `Vec<u8>`)
 - [ ] P4 economic tier (measured cost → `Materialization`) — instrument done (#5431); remaining = (a) space arm: per-shard RSS plumbing → `CostEstimate.space` on `Runnable` (nodes A+B, [plan](docs/plans/resource-aware-scheduler.md)); (b) time arm: consumer feedback → `CostAccount.time = Measured` feeds width-fold (node C+D); dissolves `cost_account_predicted_zero()` on the floor path
 - [ ] P5 native `content(T) = content_hash(subgraph)` — gated on B2
 

dsl/gunbc/roadmap_authority.dag

@@ -166,7 +166,7 @@ fn section_1() -> RoadmapSection {
           derivable_doc(id: "1-measure-foundation", prs: [5470, 5478], title: "**Section 1 spawn-width foundation — std.measure expressibility**", description: "— the FLOOR family (measure_scale_fraction_floor + measure_fit_count_floor) and the demand-CEIL family completing the ceil-not-equal-floor money-pair landed, dissolving the measure unwrap→raw-arith→rewrap §3 fork. Authority: the expressibility-frontier spec #5467", path: "docs/plans/expressibility-frontier.md"),
           authored(id: "1-placement-inert", done: false, content: "**scheduling: placement + materialization inert** — `Placement`/`Materialization` modeled + witness-passing, **no live consumer** (same band as the host-ops gap, substrate side)"),
           authored(id: "1-sched-resource-aware", done: false, content: "**scheduling: resource-aware (larger allocation → larger throughput)** — `spawn_width` is a side-channel derived from a static data row (`gunbc_ci_floor_measured_peak`, coarse whole_run÷concurrency estimate); goal: width derived from per-`Runnable` cost in the plan itself, self-calibrating from measured per-shard RSS. Four nodes: (A) measurement plumbing — claim_batch emits `[measurement] per-shard-peak-rss`, claim_executor collects max *(dispatched adhoc-97532cd3-dfe)*; (B) `CostEstimate.space` on `Runnable` in std; (C) scheduler reads cost → derives width, side-channel deleted; (D) calibration loop retires static data rows. Construction invariant: host memory increase → higher width → more throughput, no `.dag` edits required. [plan](docs/plans/resource-aware-scheduler.md)"),
-          authored(id: "1-caching-forked", done: false, content: "**caching forked** — sccache live · resolve-cache **always-on (#5789 — default `$TMPDIR/gunbc-rg-cache`, CI pending merge)** · ParseTable memo live · RecordedFixture · BuildBuddy opt-in → converge on `realize(subject)` (§2 P2)"),
+          authored(id: "1-caching-forked", done: false, content: "**caching forked** — sccache live · resolve-cache **opt-in (`GUNBC_RESOLVED_GRAPH_CACHE_DIR`; #5789 always-on default reverted — its whole-file JSON read/write OOMs the floor, must stream before it can default on)** · ParseTable memo live · RecordedFixture · BuildBuddy opt-in → converge on `realize(subject)` (§2 P2)"),
         ],
         edges: [],
       ),
@@ -241,7 +241,7 @@ fn section_2() -> RoadmapSection {
           authored(id: "2-p1", done: true, content: "P1 honest keys by construction — warm==cold purity oracle **LANDED (#5429)**"),
           authored(id: "2-p2", done: false, content: "P2 one door: `realize(subject)` sole API — kernel inhabits `cache_interface.dag` (#5446); ParseTable dissolution is downstream of the dsl→v2 de-fork (§5)"),
           plain(id: "2-hermetic", content: "hermetic fixtures feed P2: [x] M4.1 universal hermetic corpus governance (#5236, [plan](docs/plans/m4-universal-hermetic-corpus.md)); [ ] M5 fixture-store onto one Realization kernel ([plan](docs/plans/m5-fixture-store-consolidation.md))"),
-          authored(id: "2-p3", done: true, content: "P3 **resolve-cache enable** — purity proven (616/616); #5789 makes cache **always-on** (default `$TMPDIR/gunbc-rg-cache`, no env var required; CI pending merge 2026-06-25) ← **core ask landed**"),
+          authored(id: "2-p3", done: false, content: "P3 **resolve-cache enable** — purity proven (616/616); cache is correct but **opt-in** (`GUNBC_RESOLVED_GRAPH_CACHE_DIR`): #5789's always-on `$TMPDIR` default was reverted because the whole-file JSON read (hit) and write (miss) buffer ~11x the packed graph in memory and OOM the concurrent-shard floor. Re-enabling by default is gated on a streaming IO realization (binary + `deserialize_from`/`serialize_into`, no whole-file `Vec<u8>`)"),
           authored(id: "2-p4", done: false, content: "P4 economic tier (measured cost → `Materialization`) — instrument done (#5431); remaining = (a) space arm: per-shard RSS plumbing → `CostEstimate.space` on `Runnable` (nodes A+B, [plan](docs/plans/resource-aware-scheduler.md)); (b) time arm: consumer feedback → `CostAccount.time = Measured` feeds width-fold (node C+D); dissolves `cost_account_predicted_zero()` on the floor path"),
           authored(id: "2-p5", done: false, content: "P5 native `content(T) = content_hash(subgraph)` — gated on B2"),
           plain(id: "2-blockers", content: "blockers: [ ] B1 #5295 generic-instantiation (gates cross-shard `Share`) · [ ] B2 cross-tree content-hash (gates P5)"),

src/v1/stage0/src/cli_run.rs

@@ -28,8 +28,8 @@ use crate::v1_std_core::{
 use serde::Serialize;
 
 use crate::resolved_graph_cache::{
-    lookup as cross_process_lookup, resolved_graph_cache_root, subject_digest_for_closure,
-    write as cross_process_write, CacheLookupResult,
+    lookup as cross_process_lookup, resolved_graph_cache_root_from_env,
+    subject_digest_for_closure, write as cross_process_write, CacheLookupResult,
 };
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
@@ -404,8 +404,7 @@ fn resolve_entry_with_parse_cache(
 > {
     let sources = load_sources_for_entry_with_index(&index.source_files, entry_file)?;
 
-    {
-        let cache_root = resolved_graph_cache_root();
+    if let Some(cache_root) = resolved_graph_cache_root_from_env() {
         let subject = subject_digest_for_closure(&sources);
         match cross_process_lookup(&cache_root, &subject) {
             CacheLookupResult::Hit(hit) => {
@@ -515,8 +514,7 @@ fn resolve_entry_with_parse_cache(
         return Err(format_error_nodes(&ownership_diags, &source_indices));
     }
 
-    {
-        let cache_root = resolved_graph_cache_root();
+    if let Some(cache_root) = resolved_graph_cache_root_from_env() {
         let subject = subject_digest_for_closure(&sources);
         let _ = cross_process_write(&cache_root, &subject, &typed, source_indices.as_ref());
     }

src/v1/stage0/src/resolved_graph_cache.rs

@@ -65,15 +65,21 @@ pub enum CacheWriteOutcome {
     AlreadyExists,
 }
 
-pub fn resolved_graph_cache_root() -> PathBuf {
-    std::env::var_os("GUNBC_RESOLVED_GRAPH_CACHE_DIR")
-        .map(PathBuf::from)
-        .unwrap_or_else(|| {
-            let user_tag = std::env::var("USER")
-                .or_else(|_| std::env::var("USERNAME"))
-                .unwrap_or_else(|_| "shared".to_string());
-            std::env::temp_dir().join(format!("gunbc-rg-cache-{user_tag}"))
-        })
+/// The cross-process resolved-graph disk cache is **opt-in**: it activates only
+/// when `GUNBC_RESOLVED_GRAPH_CACHE_DIR` names a directory. With it unset the
+/// cache is entirely off (`None` — no read and no write).
+///
+/// Why opt-in and not a `temp_dir()` default: under the CI floor the cache is a
+/// net loss. Every commit changes the compiler binary, which is part of the
+/// content digest, so the floor re-colds each run and is overwhelmingly a miss;
+/// and both paths buffer a whole cache file in memory (a hit `read_to_end`s the
+/// verbose-JSON file, ~11x the packed graph; a miss `to_vec`s the whole JSON),
+/// so multi-GiB entries resolved across concurrent shards OOM the runner. The
+/// cache pays its worst cost exactly where it collects ~no benefit. (#5789 made
+/// it always-on via a `temp_dir()` default; this reverts to opt-in — its IO
+/// realization must stream, not buffer, before it is safe to default on.)
+pub fn resolved_graph_cache_root_from_env() -> Option<PathBuf> {
+    std::env::var_os("GUNBC_RESOLVED_GRAPH_CACHE_DIR").map(PathBuf::from)
 }
 
 fn extract_module_path(content: &str) -> Option<String> {