Skip to content

feat(egl): environment-optimized dynamic libraries wave — Codex gpu-e4-e5 carried + packages 2/4/5/7/8, U2/X2/V2, rt_* sites routed, Rust twins - #675

Closed
ormastes wants to merge 5 commits into
mainfrom
work/egl-wave-b-2026-09-12
Closed

ormastes wants to merge 5 commits into
mainfrom
work/egl-wave-b-2026-09-12

Conversation

@ormastes

@ormastes ormastes commented Sep 12, 2026 •

Copy link
Copy Markdown
Owner

What this is

The EGL wave (Environment-optimized dynamic Libraries) landing as one PR. It carries
two lanes at once:

  • Codex's codex/gpu-e4-e5-production-sol-20260912 — 24 commits (EGL packages 1,
    6 and 8 core: startup admission and variant lifecycle, GPU provider/completion
    authorities, environment policy collection and native-worker handoff), merged onto a
    gate-synced main at bd8df49e8d4. That Codex session hit its usage limit, so its
    branch was not going to land on its own; this PR carries it rather than letting the
    work rot on a branch.
  • The Claude agent wave on top of it — agents P, W, V, U, X, T, N, K plus the
    collected-receipts commit, and the U2/X2/V2 follow-ups applied from their own
    worktrees and committed here with their receipts' numbers.

49 commits ahead of main. origin/main was merged in twice (it advanced 105 then 37 commits while this ran); the first merge had one conflict, the second none.

Per package

Pkg What this branch delivers What stays with its owner
1 Codex: bounded contracts, catalog, exact admission, composite MAX→PREFER/REQUIRE selection, rejection-preserving canonical publication, generation-checked runtime commit Production startup/CLI consumer wiring, qualified self-host execution, host matrix
2 U + U2: callable_map_v1 — the loader owner every upstream module deferred to — with ten stable refusal codes; artifact bytes hashed before and after the mapping so a rename-swap inside the hash→load window is refused (SwappedDuringLoad, code 11, distinct from the pre-load mismatch because one refuses an inert artifact and the other refuses after constructors ran); operation family widened from i64(i64,i64) to i64(i64...) arity 0..6 with executed native==static parity at 0/1/2/3/6; session_callable_adapter_v1 binding a session's facets through a typed capability token that cannot be replayed across sibling sessions of the same published plan SMF placement is still unexecuted (smf_dlsym returns a section offset, exec_mapping_map_rx has no entry ABI, src/os/smf/smf_dynlib.spl is Codex's egl-prod-integration lane); artifact identity is not fd-pinned — three shared-descriptor routes were measured unavailable on this seed/kernel, including /proc/self/map_files returning EPERM; the one-statement production hook is stated, not made, because every candidate call site is Codex-owned
4 V + V2: parse_result_provider_seam_v1 over a frozen five-source corpus (a struct of callables, no registry); dialect identity computed from the tables a dialect executes rather than declared — four independent digests (grammar / actions / schema / semantic) whose probe domain is derived from the tables themselves, so an entry that changes moves the digest with nothing else edited; SDN admitted as a second dialect through the same seam; a compatibility example proving the seam result is byte-identical to what the frontend's public entry already produces The seam is still not called by parse_full_frontend_with_scope — frontend.spl is Codex-owned, and the hook must not be applied until per-request identity verification (~2 s/request on this seed) is replaced by an admit-once entry; sosh is blocked (below)
5 W + P: nine published candidate classes derived two independent ways; scalar state resolution through sealed and unsealed rows; issue/verify routing receipts where verification recomputes masks and resolved state from the claimed bytes and trusts no receipt field; an advisory frontend wrapper; boundary-state qualification and the AVX2 classifier primitive with its contract spec Native x86 / self-host qualification — this host is aarch64 (Cortex-X925, no AVX2), so the AVX2 primitive is contract-verified, not executed; broader lexical coverage; emitted v3/v4 siblings
6 Codex: GPU admission and Engine2D completion bridges, hosted ABI-v1 mapper/lifetime owner, bounded capability tokens, retained unload fences. T adds the packed completion authority route on the queue owner with a single-use redemption ledger Runtime-owned exact device-program image token, Vulkan parser kernel, real device fixture
7 X + X2: a real admit → resident → retire island with transfer-aware scheduling; a deferral feedback loop where one round's deferred set feeds the next, proven over a live loop with a task whose cost exceeds the whole budget (forced at round 4 behind 3 deferrals, read from the loop rather than restated); a device execution body joined to T's authority route, so the CPU rasterizer refuses a device-lane lease (LaneMismatch) and a spent authority stays spent even when the host queue then disagrees; measured transfer costs from the backend's own damage/readback receipts instead of declared numbers, with the sticky-damage limitation pinned as a checked number (4096 → 8192 bytes) rather than docstring prose Authority issuance from a real Vulkan fence is Codex-fenced — nothing in this tree can tell a real fence from a fabricated record; device-lane measured bytes return zeros because the island moves none itself; incremental (non-sticky) upload measurement needs per-crossing damage clearing, which only present() owns
8 Codex: pure policy resolver, presence-aware supplied-layer collector, authenticated sealed collect-to-prepare boundary, catalog projection, preparation bundle. N adds a host-capability ledger probe (scripts/audit/egl-host-capability-ledger.shs, 444 lines) that describes a host in four explicit states and never blocks a push — deliberately an audit, not a gate Trusted command/config owner wiring, package integration, explain output, measured NFR evidence — all Codex's; this branch adds no consumer
— K: 116 orphaned guards wired into the repo-hygiene advisory job, and the corresponding 118 rows removed from the unwired baseline

Gate verdicts

Binary for every spec number below: /home/yoon/dev/simple/bin/release/aarch64-unknown-linux-gnu/simple,
50093192 bytes, 2026-09-06 09:59:11 +0900, sha256 3d120a6f9ab5704b… — the deployed
Rust seed. No rebuild: the changes are src/lib and src/compiler source, read on
every run.

Gate Verdict
check-no-direct-rt.shs --roots src --rev HEAD --baseline-rev origin/main PASS — 16633 file(s) scanned (roots=src, src=6335), forbidden=6335, extern_decls=6640 (base origin/main: 6335) — was 6344, i.e. +9
check-guard-wiring.shs PASS — 1686 guard(s) checked, 642 invoked, 1023 orphaned (607 baselined), 0 NEW unwired, 0 stale, 0 copied hook(s)
check-workspace-root-guard.shs OK
check-test-tree-divergence-delta.shs origin/main HEAD PASS — 3209 pre-existing offender(s), 0 introduced by this range (base is itself FAIL — 3943 diverged vs 965 baselined; list recorded, below)
check-c-runtime-compiles-push.shs PASS — 144 file(s) compiled, 0 errors (5 skipped for unavailable external deps)
check-runtime-api-regression-push.shs origin/main..HEAD PASS — 3172 symbol(s) checked, 0 removed
check-merge-content-conservation-push.shs origin/main..HEAD PASS — 48 commit(s), 3 merge(s), 203 merged path(s), 0 add/add drops, 0 lost definitions (selftest 6/6)
check-cpu-hotloop-idiom.shs cpu_lane_hotloop_ok=true
check-runtime-source-list-parity.shs PASS — 149 file(s) checked, 0 drift
cargo check --release --bin simple (CARGO_TARGET_DIR=/home/yoon/cargo-egl) Finished, 0 errors, 7 warnings
check-bug-status-consistency.shs FAIL — 8 disagreements, all 8 pre-existing on origin/main (verified per file); this branch's two new records add none
check-push-must-pass.shs --self-test FAIL — valid fixture rejected, pre-existing on origin/main: its three inputs are byte-identical here, and the same command in a detached worktree at origin/main (202b68d) answers identically

Gate work done on the branch, not deferred

  • The +9 direct rt_* sites that refused the earlier push are gone. They were
    found by dumping --offenders at both revisions and diffing per file: 8 in
    host_environment_snapshot_v1.spl (new, Codex) and 1 in simd_capabilities.spl.
    Fixed by adding four typed aliases (cpuid, xgetbv, cpu_is_x86_64,
    x86_avx512_os_state_usable) to src/lib/nogc_sync_mut/sffi/host.spl — the
    allowlisted provider whose header says product code should import from it instead of
    declaring extern fn rt_* — and deleting Codex's dead host_x86_avx512_os_state_usable_native
    wrapper (zero callers in src/ or test/). The baseline file was not touched and
    the allowlist was not extended.
    Two "fixes" were considered and rejected in the
    commit message: reimplementing the C probe in Simple (loses the
    SIMPLE_RUNTIME_FORCE_NO_X86_XSTATE test override and the non-x86 false), and
    allowlisting simd_capabilities.spl (would retire 13 pre-existing violations as a
    side effect — a policy change for the guard's owner).
  • Two NEW unwired guards the merge exposed, each handled by its own evidence:
    check-gpu-provider-abi-lifetime-v1.shs (Codex's, measured rc=0 with a real verdict
    line) is wired into the advisory job; check-test-runner-executes-bodies.shs
    (arrived from main already unwired there, executes bin/simple, honestly RED here)
    is baselined with the measurement written next to it — wiring it would install a
    red gate and an opt-out line would be a false statement.
  • One manifest row the merge silently dropped — chrome-layout-geometry-diff in
    config/check/must_check_gates.sdn, present on main, absent in both the merge base
    and the branch tip, gone from the merge result. Restored verbatim; that file is now
    byte-identical to main. This is the stale-merge content-loss class vcs.md warns
    about: no conflict, no failing check.
  • A false positive in check-merge-content-conservation-push.shs fixed at the
    source: its definition extractor matched ^(pub )?fn NAME, so when one parent had
    fn shared_text_is_ascii and main widened it to pub(crate) fn, the correctly
    merged result looked like a lost definition. Widened to accept a Rust visibility
    scope — which makes the extractor see more definitions on every side, so it cannot
    mask a real loss. Selftest still 6/6, including the lost-definition incident replay.

Specs

36 specs added or modified by this branch were run individually; 31 green.
Every Claude-agent and U2/X2/V2 spec passes, including the three that gate the work
above: host_environment_snapshot_v1_spec 12/12 (the module whose extern fn rt_*
declarations were replaced), session_callable_adapter_v1_spec 9/9,
parse_dialect_identity_and_adapters_v1_spec 11/11,
environment_resident_executor_deferral_v1_spec 8/8,
parse_result_provider_seam_v1_spec 8/8 (unchanged, no example edited).

4 are red, all Codex's, and all red at Codex's own base. Verified by running them
in a detached worktree at bd8df49e8d4 on the same binary — the four verdict lines
come back byte-identical (5/8, parse-error with 0 executed, 3/6, 4/5). Filed as
doc/08_tracking/bug/egl_gpu_e4_e5_base_four_red_specs_2026-09-12.md with the
attribution evidence rather than fixed here: they are the environment policy/handoff
surface the Codex lane owns end to end, that session is out of budget, and guessing at
another lane's semantics is how a carrier becomes a clobber.

The whole test/01_unit/compiler/driver/ directory was additionally swept per file
for the parse-seam work — 182 specs at HEAD and again at Codex's base bd8df49e8d4.
179 of 182 are byte-identical base→HEAD. The three that differ are benign: one
cost spec times out on both sides, and V's and V2's two new specs are absent at base
and green at HEAD (8/8 and 11/11). That directory carries a large pre-existing red
backlog (85 of 182 non-OK on both sides), which is why it is reported as a delta.

Known blocked, stated rather than hidden

  • Agent Q did not land. src/compiler/80.driver/parser_variant_build_plan_v1.spl:89-92
    reassigns a local named feature, which the deployed compiler parses as the spec-DSL
    group keyword, so the whole module fails with
    parse: Unexpected token: expected Colon, found Assign and every consumer of
    ParserVariantBuildPlanV1 is blocked. The fix is a two-token rename
    (feature → feature_id); it is the only occurrence of the pattern in src/ and it
    belongs to that module's Codex owner, so it is stated here rather than applied.
  • Agent S is frozen — no receipt was produced for it, so nothing from that agent is in this branch.
  • sosh dialect adapter (package 4) — os.apps.shell.shell_script.StmtKind and the
    frontend's StmtKind collide, and enum-variant resolution is program-global by name,
    so the frontend's wins inside shell_script.spl itself; an aliased import cannot
    help. The adapter was written, measured red, and deleted rather than left as dead or
    skipped code. doc/08_tracking/bug/shell_stmtkind_collides_with_frontend_stmtkind_2026-09-12.md.
  • Frontend rewiring (package 4) — the seam is not called by
    parse_full_frontend_with_scope; frontend.spl is Codex-owned and the hook is gated
    on the ~2 s/request identity-verification cost being replaced by an admit-once entry.
  • Pre-existing test-tree divergence stepped over on a delta-PASS; the offender list
    is recorded at doc/08_tracking/bug/test_tree_divergence_preexisting_egl_wave_2026-09-12.{md,txt}
    as vcs.md requires.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RWALUeGMTW8V9QKacRoE1r

ormastes and others added 2 commits September 12, 2026 21:44
…4-e5 (24 commits) carried + packages 2/4/5/7/8 non-Codex halves, follow-ups U2/X2/V2, direct rt_* sites routed, Rust twins for 8 C-only rt_*

Single commit rooted on origin/main (the branch history of PR #626 is superseded: a merge of main into a pushed
branch is refused by the 64-commit outgoing bound and a squashed merge re-conflicts on every later main edit).
Content = the wave branch merged with main; provenance, per-package table and gate verdicts are in the PR body.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RWALUeGMTW8V9QKacRoE1r
…m (mcp-stdio-roundtrip t32-mcp-server-runnable )

The push gate validate_ledger_text requires one must_check_db.sdn row per bootstrap-tier
must_check_gates.sdn row; these ids landed on main with manifest rows only. Rows added as
todo/never with the manifest command byte-identical.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RWALUeGMTW8V9QKacRoE1r
ormastes and others added 3 commits September 12, 2026 23:02
…tal.shs (its opt-out line was stale on this tree)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RWALUeGMTW8V9QKacRoE1r
…ard) into the advisory step

check-guard-wiring.shs --rev reported it as NEW unwired on the pushed tree.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RWALUeGMTW8V9QKacRoE1r
@ormastes ormastes closed this Sep 12, 2026
@ormastes ormastes reopened this Sep 12, 2026
@ormastes
ormastes temporarily deployed to self-review-admission September 12, 2026 15:14 — with GitHub Actions Inactive
@ormastes

Copy link
Copy Markdown
Owner Author

Superseded by #693 (same content re-rooted as one commit on current main).

@ormastes ormastes closed this Sep 12, 2026
@ormastes
ormastes deleted the work/egl-wave-b-2026-09-12 branch September 12, 2026 15:19
@ormastes
ormastes temporarily deployed to self-review-admission September 12, 2026 15:24 — with GitHub Actions Inactive

This branch was previously deployed

1 inactive deployment
self-review-admission — 04ea3d5f Deployed Sep 12, 2026 by ormastes via Invalidate changed PR self-review state #4361
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant