Skip to content

fix(m2): address independent audit findings (P1, P2, safe P3) - #15

Merged
brasga-a merged 11 commits into
masterfrom
fix/m2-audit-findings
Nov 18, 2026
Merged

brasga-a merged 11 commits into
masterfrom
fix/m2-audit-findings

Conversation

@brasga-a

Copy link
Copy Markdown
Owner

Fixes from the independent Milestone 2 audit. Stacked on #14; one commit per finding.

Changes

  • P1, F3D-INV-12 (602b049): cooked-shader reads are bounded before allocation (read_cooked_shader_bounded). Non-regular files and files over MAX_COOKED_SHADER_BYTES fail with ENGN-0008 based on file metadata, before any content is read. A 512 MiB artifact previously pushed peak RSS to 558 MiB; it now fails at 83 MiB.
  • P2-1, SPN-INV-001 (b2fc73a): FND-T09 and F3D-T13 now take dependency edges from Cargo's resolver (cargo tree --invert, normal + build, --target all). Previously a hand-written TOML parser built them and missed three valid manifest forms: [dependencies.X] tables, renamed deps, and [build-dependencies].
  • P2-2, F3D-INV-08 (8ee9968): recorded golden to_bits() values for the camera helpers (F3D-T04) and for 600-tick camera poses (F3D-T05c). Windows CI now checks bit identity across platforms, not just consistency within one process.
  • P2-3, Gate G2 (1011c74): new nightly fuzz job runs cargo fuzz run --sanitizer address for 300 s over a generated seed corpus, as an owner-approved ADR-0007 exception that applies to this job only. libfuzzer-sys is registered in DEPENDENCIES.md, NCSA is allowed only for that crate, and the uninstrumented-run claim is withdrawn from VERIFICATION.md.
  • Safe P3 (2194833):
    • RendererInfo::sync_validation_enabled is exposed, and Validation::Required now enforces it.
    • retire_buffer/retire_image no longer leak handles when given a rejected suballocation.
    • reclaim_completed uses retain.
    • Allocator shutdown no longer calls debug_assert! from Drop.
    • Corrected opt-in GPU counts in the docs.

Verification (local, reference RX 460 / Hyprland)

  • ./scripts/check.sh: exit 0. CARGO_TERM_COLOR=always cargo test --workspace: green.
  • I added each of the three missed dependency forms to crates/spawn-ecs/Cargo.toml in turn. FND-T09 and F3D-T13 failed for each one; the change was reverted afterwards.
  • Opt-in GPU tests: 10/10 pass with VK_LAYER_KHRONOS_validation and sync validation, 0 errors. Opt-in slangc 2026.22 test: 1/1.
  • 2 s auto-closing smoke run with validation: clean exit, validation_errors=0, no leaks at shutdown.

Not verified locally

  • The fuzz job: nightly and cargo-fuzz are not installed on this host. It runs for the first time on this PR. Its log should show cov: > 0, not corp: 1/1b.
  • The golden bits were recorded on Linux. The Windows job on this PR is the cross-platform check.

Still owner-only for Gate G3

  • Interactive verification protocol.
  • Performance budget decision. About 53 fps was observed on a 60 Hz FIFO display; this needs per-frame timing before drawing any conclusion.
  • master branch protection and default-branch alignment.

🤖 Generated with Claude Code

brasga-a and others added 11 commits November 18, 2026 15:01
…V-12)

`LinuxBackend::read_cooked_shader` used `fs::read`, loading the whole file
before `CookedShader::decode_named` enforced `MAX_COOKED_SHADER_BYTES`. A
512 MiB artifact raised peak RSS from 83 MiB to 558 MiB, and a symlink to a
character device such as `/dev/zero` would grow without bound.

Add `read_cooked_shader_bounded` in `src/options.rs` (always compiled, so it
is tested headlessly on Linux and Windows): it rejects non-regular files and
files whose metadata size exceeds the cap with `ENGN-0008`, then reads at
most `MAX + 1` bytes so files that grow during the read are still rejected.

Verified: unit tests (sparse MAX+1, exact cap, directory, missing file,
`/dev/zero`); 512 MiB probe now fails with ENGN-0008 at 83 MiB peak RSS.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…V-001)

FND-T09 and F3D-T13 derived `spawn-*` edges from a hand-written TOML line
parser that silently missed valid Cargo forms: `[dependencies.<name>]`
tables, renamed dependencies (`alias = { package = "<name>" }`), and
`[build-dependencies]`. An engine crate could depend upward on
`spawn-engine` or a tool without either test failing.

Both tests now share `tests/common/graph.rs`, which takes direct edges from
Cargo's resolver (`cargo tree --workspace --invert <pkg> -e normal,build
--target all --depth 1`), so every manifest form and every target is
covered. The duplicated parsers and acyclicity helpers are removed, and
F3D-T13 now also asserts no upward edges.

Verified: new synthetic regression test; injecting each of the three forms
into crates/spawn-ecs/Cargo.toml makes FND-T09 and F3D-T13 fail (reverted).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
… (F3D-INV-08)

F3D-T05 compared camera poses only within one process (cadence
equivalence), and F3D-T04 checked `quat_from_yaw_pitch_f64` and
`yaw_pitch_basis_f64` only with tolerances, so running on Windows CI proved
self-consistency, not the required Linux/Windows bit identity.

- F3D-T04: recorded `to_bits()` table for the camera orientation/basis
  helpers, including pole-limit and near-±π inputs.
- F3D-T05c: 600-tick fixed input schedule (scripted ticks 0..48 plus an
  integer-derived look/move pattern) with recorded pose bits at ticks
  1, 60, 300 and 600. Mismatch messages print a re-recordable table.

Values recorded on Linux x86_64; identical in debug and release builds
locally. Windows CI confirms the same bits on this PR.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ser (Gate G2)

The only fuzzing evidence for `decode_cooked_shader` was an uninstrumented
stable libFuzzer run: without SanitizerCoverage it reports "no interesting
inputs" and the corpus collapses to `1/1b`, so it exercised almost nothing.

- CI job `fuzz`: installs nightly only for this job (owner-approved
  ADR-0007 exception), generates a synthetic seed corpus, runs
  `cargo fuzz run --sanitizer address decode_cooked_shader` for 300 s,
  uploads crash artifacts on failure, and runs cargo-deny on the fuzz
  package.
- `crates/spawn-render/fuzz/examples/seed_corpus.rs`: writes raw and
  structured-path seeds (vertex/fragment/uniform/compute) that round-trip
  through `CookedShader`; seeds are generated, never committed.
- `libfuzzer-sys` registered in DEPENDENCIES.md; deny.toml allows NCSA
  (bundled LLVM libFuzzer, permissive per ADR-0008) for that crate only.
- ADR-0007, DECISIONS.md, GATES.md (G2 requires instrumented runs) and
  spawn-render AGENTS.md record the exception; VERIFICATION.md withdraws
  the uninstrumented claim and corrects the opt-in GPU count to 9/9.

Not verified locally: nightly/cargo-fuzz are not installed on this host;
the job runs for the first time on this pull request. cargo-deny passes
locally for both the workspace and the fuzz package.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
- `RendererInfo::sync_validation_enabled` exposes whether Khronos
  synchronization validation (`VK_EXT_validation_features`) is active;
  `Validation::Required` now fails with REND-0008 unless debug-utils and
  sync validation are both available, and F3D-T11/F3D-T12 assert it.
  Previously sync validation was silently optional.
- `retire_buffer`/`retire_image` queue the owned Vulkan handle before
  validating the suballocation, so a stale or forged suballocation returns
  an error without leaking the VkBuffer/VkImage (new GPU regression test).
- `reclaim_completed` uses `retain` instead of building a new `Vec` on
  every call.
- Allocator shutdown no longer `debug_assert!`s on leaks (it runs from
  `Drop`); leaks stay visible in `stats().live_allocations` and tests
  still require zero.

Verified: opt-in GPU suite 10/10 on RX 460 with validation + sync
validation, 0 errors.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…mance G3 item

The CI `fuzz` job on PR #15 ran `cargo fuzz run --sanitizer address
decode_cooked_shader` for 300 s: 34,658,770 executions, coverage 328 -> 630,
corpus 5 -> 155 units, 0 crashes/leaks/OOM, cargo-deny clean for the fuzz
package. Exit-gate item 3 is now Verified with a link to the run.

Gate G3 stays IN PROGRESS. The open owner items are the interactive
verification protocol and the performance-budget decision (the milestone
defines no numeric budget). The measured ~53 fps on a 60 Hz FIFO display is
recorded as an observation still to be explained.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Three live runs by the owner on the reference RX 460 / Hyprland host,
with Khronos validation and sync validation enabled:
- WASD and vertical translation; diagonal speed stays at or below ~6 m/s.
- Mouse look only while the cursor is locked; yaw wraps across ±π.
- Pitch clamps at ±1.57 rad without the image flipping.
- Focus loss releases the cursor lock and clears held movement.
- Live resize 769x504 -> 2530x1024 recreates the depth attachment.
- `Escape` exits cleanly; validation_errors=0; no leaked GPU allocations.

The owner visually confirmed the orientation, occlusion, pole clamp, and
that the resize showed no corruption.

Exit-gate item 5 is now Verified. Gate G3 stays IN PROGRESS, pending only
the owner performance-budget decision and final sign-off. Interactive
mouse-drag resize and live 0x0/minimize suspension remain listed as not
exercised.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…lose run

Fourth owner run: about 55 continuous drag-resize events (769x504 to
1468x816 and back) with Khronos validation and sync validation. The depth
attachment was recreated and released on each event, the camera pose was
unchanged, and the owner confirmed no corruption, flicker, or aspect
distortion. The compositor close exited cleanly (CloseRequested) with
validation_errors=0 and no leaked GPU allocations.

The catch-up clamp dropped 6 single ticks (100 ms total) during the drag:
some resize frames took longer than ~83 ms because each event triggers
vkDeviceWaitIdle plus swapchain and depth recreation, in a debug build
with validation. Tick stayed monotonic. Recorded as an observation.

Mouse-drag resize is removed from the not-exercised list.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
`choose_image_count` requested `minImageCount + 1` images. RADV on
Wayland reports `minImageCount = 3`, so the swapchain had 4 images. In
FIFO mode the client could then queue 4 frames ahead of the display.

Measured on the reference RX 460 / Hyprland at 60 Hz, release build,
using a probe with wp_presentation feedback on the real engine path:
- submit -> scanout: 66.7 ms -> 50.0 ms (one fewer queued refresh)
- input arrival -> first visible change p50/p95: 90/99 ms -> 73/82 ms
  (with MESA_VK_WSI_DEBUG=nowlts)
- stall in the first frame after a resize: 80 ms -> 63 ms
- pacing unchanged: 60.1 fps, p99 frame interval 16.84 ms

New policy: `max(minImageCount, 3)`, capped by the surface maximum.
Surfaces with a minimum of 2 keep triple buffering; larger minimums are
honored.

Verified: unit test updated; opt-in GPU suite 10/10 with validation and
sync validation; scripts/check.sh; 2 s smoke with validation_errors=0.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
…e measurements

Adds "Performance budget and measurements (reference hardware)" to the
first-3d VERIFICATION.md. The owner approved the Milestone 2 local budget
on 2026-11-18; all release-build measurements on the RX 460 / Hyprland
reference host pass:
- pacing: 60.1 fps, p99 16.84 ms, with MESA_VK_WSI_DEBUG=nowlts
- CPU per frame 0.10 ms; 0.35 us per draw; simulation 0.003 ms per tick
- 20 MiB GPU memory blocks; 75-83 MiB RSS
- swapchain and depth recreation frame 3.9 ms
- input arrival to first visible change at scanout: p95 82 ms
  (budget <= 100 ms), measured with wp_presentation feedback

Also records:
- the latency breakdown and the Mailbox lower bound (35 ms p95);
- the next-milestone target (p95 <= 50 ms via VK_KHR_present_wait pacing
  and late latching);
- the 63 ms stall after a resize;
- the Mesa 26.2.5 / Hyprland 0.56 commit-timing pacing issue, reproduced
  with vkcube, and its workaround.

The performance item of Gate G3 is now Verified. The gate stays
IN PROGRESS pending only the owner's final sign-off and the stack merge.
PLAN.md, root AGENTS.md and spawn-render AGENTS.md are synchronized.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
The owner reviewed the exit-gate evidence and signed off on 2026-11-18:
automated tests, GPU tests with validation and sync validation,
coverage-guided ASan fuzzing, the four interactive runs, and the
approved performance budget. Milestone 2 Phases 0-6 are EXISTS on
Linux/Wayland + Vulkan 1.3. X11 and Windows remain PLANNED, and the
unexercised edges stay listed.

The status moves from "IN PROGRESS - PENDING OWNER VERIFICATION &
SIGN-OFF" to "SIGNED OFF BY OWNER (2026-11-18)" across docs/first-3d/*,
docs/README.md and the root AGENTS.md. F3D-T13 now requires the sign-off
wording and rejects the stale pending status.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
@brasga-a
brasga-a changed the base branch from feat/m2-p5-engine-camera to master November 18, 2026 19:40
@brasga-a
brasga-a merged commit 5cac4fd into master Nov 18, 2026
4 checks passed
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