conceit v0.1.0

First public release. conceit began as the private build orchestration behind a
working PyTorch 2.14.0a0 + vLLM stack on Blackwell sm_120 — CUDA 13.3, Python
3.14, gcc-15, both built from source with torchvision and torchaudio compiled
against the local torch wheel. Everything below is what it took to turn that
into something another person can run.

Added

• scripts/build-upstream.sh — the orchestrator. Clones, syncs submodules
  with exponential backoff, runs a preflight (kills stale builds by process
  group, reports memory/disk/cpu, surfaces the target repo's build
  instructions), creates the venv, compiles under a recursive process-tree
  monitor, and produces wheels. Presets carry the per-target quirks: vLLM's uv
  invocation with a local-torch override file that keeps cuda-toolkit out of
  the dependency graph, PyTorch's --no-build-isolation, llama.cpp's cmake
  flags.
• .build-status.json — machine-readable build state per target, written
  atomically so a reader never catches a half-written file. Phases run
  init → clone → venv → prebuild → build_start → building → build_done → wheel → done, carrying the current ninja step and a live count of every
  descendant process. Meant to be polled instead of grepping a log.
• scripts/cuda-env.sh — one place for every default: CUDA_HOME,
  TORCH_CUDA_ARCH_LIST=12.0+PTX, gcc-15 as the CUDA host compiler, and the
  cuDNN/cuBLASMp/cuDSS staging paths. Every value is ${VAR:-default}, so any
  single knob is overridable without editing the file.
• scripts/gen-patches.sh — exports the live diffs under src/ into
  patches/, reporting the upstream ref each one was cut from.
• scripts/install-cuda-toolkit.sh — runs NVIDIA's runfile installer
  behind a temporary gcc-15 symlink directory, without touching the system
  compiler.
• patches/pytorch/local.patch — cuDSS 0.8.0 changed
  cudssMatrixCreateCsr to take three separate type parameters and split its
  type enum away from cudaDataType_t. Verified against pytorch d9abf9e1053.
• patches/vllm-deps/vllm-flash-attn.patch — vllm-flash-attn gates on
  Python 3.9–3.13, so a 3.14 build fails at cmake configure. Verified against
  vllm-flash-attn b3964b1.
• .agents/skills/build-triage/SKILL.md — an abductive-triage
  instantiation for build failures: evidence hierarchy, coordinates-before-
  compilation, and a Step-2 adversarial table that scores each assumption
  against the simplest alternative that makes it wrong. Every row in that table
  cost a real compile.
• Makefile — the interface. Core builds, the torchvision/torchaudio matrix
  across both venvs in full-CUDA and python-only variants, patch generation and
  replay, smoke tests, check-env, and the quality gates.

Fixed

• The progress monitor watched the wrong process for every build ever run.
  eval "$cmd" 2>&1 | tee log & sets $! to tee's PID, not the build's, so
  count_descendants reported zero active processes and the liveness loop
  tracked tee's lifetime instead of the compile's. Measured before the fix:
  ps -p $build_pid -o comm= returns tee, descendants 0. After: the real
  worker, descendants non-zero. The build now redirects to its log with a
  separate backgrounded tail. (Exit codes were never wrong here — pipefail
  was already returning the build's status — but the monitoring, which is the
  advertised feature, was blind.)
• make patch-vllm-deps could not have worked. It applied its patch at the
  vLLM repo root, but the patch is cut from
  .deps/vllm-flash-attn-src/, which cmake FetchContent populates. Applying it
  failed on hunk 1 every time.
• Patch application reported success on failure. The apply loop printed
  WARN: may already be applied and exited 0 for any failure, including a
  patch that genuinely no longer applied — the premature-success-on-exit-0
  pattern the triage skill exists to catch. It now distinguishes applied,
  already-applied, and failed, and a failure is a hard error naming the
  regeneration command.
• Nothing worked outside one specific home directory. The Makefile and
  every script hardcoded a single absolute path, so a clone anywhere else
  silently built into the wrong tree. All paths now derive from the repo root,
  computed from each script's own location.
• MAX_JOBS was pinned to 20, which wastes a bigger machine and thrashes a
  smaller one. It derives from nproc and stays overridable.
• .build-status.json could emit invalid JSON. Build commands and repo
  URLs land in the detail field unescaped, so a quote or backslash broke the
  file exactly when something interesting was happening. Values are escaped.
• An interrupted build orphaned its helpers. The tail and monitor processes
  were only killed on the normal paths; Ctrl-C left both running. A single EXIT
  trap now reaps them.
• gen-patches.sh's fallback path produced a corrupt patch. For a source
  tree without .git, it diffed against an empty baseline, which lists every
  file in the tree as a new file. There is no way to recover an upstream
  baseline from such a tree, so it now says so and skips.

Changed

• Quality gates run over one language. install-cuda-toolkit.sh was zsh
  for no reason beyond how it was first written, which forced a shebang-sniffing
  split in the Makefile, a second linter, and a zsh install in CI. Ported to
  bash; shellcheck now covers every script and CI delegates to make check so
  it cannot drift from a local run.
• Five intentional shellcheck findings carry inline suppressions with their
  reasons — deliberate word-splitting of package lists, and quotes that must
  survive into an eval — rather than lowering the severity floor globally.
• The Makefile's eight torchvision/torchaudio targets share four canned
  recipes instead of four copies of the same body, and recipes now run under
  bash explicitly. They use source, which does not exist in dash, so every
  build target failed on a Debian or Ubuntu host.
• cuDNN, cuBLASMp, and cuDSS are located by globbing their extracted
  archives under cuda/ rather than by pinned version strings, so an upgrade
  is an extract and nothing else.

Removed

• scripts/build-{pytorch,vllm,llama}.sh — three unreferenced wrappers
  holding a second copy of the upstream repo URLs, waiting to disagree with the
  Makefile.
• patches/patch-conceit.sh — a one-shot personal migration that required
  a file never present in the tree.