chore: add issue forms, a PR template and CODEOWNERS

.github/CODEOWNERS

@@ -0,0 +1,2 @@
+# Default owners for everything in the repository.
+* @vyncint

.github/ISSUE_TEMPLATE/build_failure.yml

@@ -0,0 +1,43 @@
+name: An admitted candidate failed to build
+description: >-
+  The gate passed a configuration and `cuda-oxide` then failed to compile
+  it. This is a known hole with a known cause; reports still help, because
+  each one names another codegen-time const the gate does not evaluate.
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: >-
+        `docs/LIMITATIONS.md` ("cuda-oxide is alpha") records why this
+        happens: `cargo check` under the reconverge driver does not evaluate
+        all codegen-time consts, so a gate-clean candidate can still fail the
+        real compile — an invalid `#[unroll]` factor is the case that was
+        found first.
+  - type: input
+    id: config
+    attributes:
+      label: The configuration
+      description: "As the report prints it, e.g. `block_x=64 tile=128`."
+    validations:
+      required: true
+  - type: textarea
+    id: error
+    attributes:
+      label: The compiler error
+      render: text
+    validations:
+      required: true
+  - type: textarea
+    id: kernel
+    attributes:
+      label: The kernel and its `kernel.toml`
+      render: rust
+    validations:
+      required: true
+  - type: input
+    id: pin
+    attributes:
+      label: cuda-oxide pin
+      description: "`scripts/check-pins.sh` lists every site; rust-toolchain.toml names the set."
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,9 @@
+blank_issues_enabled: false
+contact_links:
+  - name: A wrong verdict from the convergence analyzer itself
+    url: https://github.com/vyncint/reconverge/issues/new?template=false_positive.yml
+    about: >-
+      launchbound ships reconverge as a component and does not reimplement it.
+      If the finding is wrong about the kernel rather than about the launch
+      shape, it belongs upstream — zero false positives at default confidence
+      is a release requirement there.

.github/ISSUE_TEMPLATE/measurement.yml

@@ -0,0 +1,48 @@
+name: A measurement looks wrong
+description: >-
+  A timing, a ranking, or a statistic does not match what you observe —
+  including the model's estimate disagreeing with silicon.
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: >-
+        Before filing: a result is valid only for the GPU, driver and
+        compiler in its provenance, and `sm_75` and `sm_86` do not transfer
+        (`docs/LIMITATIONS.md`, "Results do not port"). Two configurations
+        whose 95% intervals overlap are reported indistinguishable and are
+        deliberately not ranked.
+  - type: dropdown
+    id: backend
+    attributes:
+      label: Backend
+      options:
+        - cuda (measured on NVIDIA silicon)
+        - metal (measured on Apple Silicon — no convergence gate on this path)
+        - model (estimated, not measured)
+    validations:
+      required: true
+  - type: textarea
+    id: observed
+    attributes:
+      label: What you saw, and what you expected
+    validations:
+      required: true
+  - type: textarea
+    id: results
+    attributes:
+      label: The run directory
+      description: >-
+        `launchbound report <run> --json`, or the `results.json` itself. It
+        carries the device name, the driver version, the warmup and repeat
+        counts and every timing sample, which is most of what this needs.
+      render: json
+    validations:
+      required: true
+  - type: textarea
+    id: provenance
+    attributes:
+      label: Provenance
+      description: "`launchbound --version`, the cuda-oxide pin, the nightly, the GPU, the driver and CUDA versions."
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/wrong_verdict.yml

@@ -0,0 +1,70 @@
+name: The gate admitted or refused the wrong configuration
+description: >-
+  A configuration launchbound admitted is convergence-unsafe, or one it
+  refused is safe. This is the report that matters most — refusing a fast
+  configuration is the product, and getting that wrong in either direction
+  is the product failing.
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: >-
+        If the analyzer is wrong about the *kernel* rather than about the
+        *launch shape*, it belongs in
+        [reconverge](https://github.com/vyncint/reconverge/issues) — this
+        project ships it as a component and does not reimplement it.
+        `docs/SAFETY.md` has the decision rule this form is about.
+  - type: textarea
+    id: kernel
+    attributes:
+      label: The kernel
+      description: >-
+        Minimal `#[kernel]` source, plus the `kernel.toml` that declares its
+        space. A corpus kernel is fine — name it instead.
+      render: rust
+    validations:
+      required: true
+  - type: input
+    id: config
+    attributes:
+      label: The configuration
+      description: "As the report prints it, e.g. `block_x=64 tile=128`."
+    validations:
+      required: true
+  - type: dropdown
+    id: direction
+    attributes:
+      label: Which way is it wrong?
+      options:
+        - Admitted, and it is unsafe (a false negative — the worst case)
+        - Refused, and it is safe (a false positive)
+    validations:
+      required: true
+  - type: textarea
+    id: evidence
+    attributes:
+      label: Why you believe that
+      description: >-
+        A hang on hardware, a `compute-sanitizer synccheck` report, a lane
+        argument, or the `cargo reconverge check --strict` output for the
+        shape. Anything is better than nothing; a hang with a driver version
+        is best.
+    validations:
+      required: true
+  - type: input
+    id: cc
+    attributes:
+      label: "`--cc`"
+      description: "The capability you ran at. A verdict at one does not transfer to another (docs/LIMITATIONS.md)."
+    validations:
+      required: true
+  - type: textarea
+    id: provenance
+    attributes:
+      label: Provenance
+      description: >-
+        `launchbound --version`, the cuda-oxide pin, the nightly, and — if a
+        measurement is involved — the GPU, driver and CUDA versions. The
+        report directory's `results.json` carries most of this.
+    validations:
+      required: true

.github/pull_request_template.md

@@ -0,0 +1,13 @@
+## Summary
+
+<!-- What does this change, and why? -->
+
+## Checklist
+
+- [ ] Title is a scoped Conventional Commit (e.g. `bench: …`), imperative, ≤ 72 chars
+- [ ] Every commit is signed off (`git commit -s`; the DCO trailer matches the author)
+- [ ] No AI attribution anywhere (trailers, message bodies, identities)
+- [ ] Tests added or updated, and the new test was watched to fail once
+- [ ] `just ci` is green locally
+- [ ] Does this move a measured number? If so the CHANGELOG entry is marked `bench:` and names the hardware
+- [ ] Does this touch the lockstep pin set (nightly / cuda-oxide / reconverge)? If so it is its own commit and `just pins` passes

CHANGELOG.md

@@ -9,6 +9,26 @@ change measured timings are marked `bench:`.
 
 ## [Unreleased]
 
+### Added
+
+- **Issue forms, a pull-request template and `CODEOWNERS`.** `.github/` held
+  `scripts/` and `workflows/` and nothing else.
+
+  The forms matter more here than they would for a library, because of what a
+  report has to contain before anyone can act on it: a result depends on the
+  GPU, the driver, the CUDA version, the compute capability, the cuda-oxide
+  commit, the nightly and which of the three backends produced it, and
+  `docs/LIMITATIONS.md` has a section on exactly that. "The tuner picked a
+  slow configuration" is unactionable; the same report with its provenance
+  block is a bug.
+
+  Three of them — a wrong convergence verdict in either direction, a
+  measurement that disagrees with what you observe, and the known hole where
+  a gate-clean candidate still fails the real compile. Blank issues are off,
+  and a finding that is wrong about the kernel rather than the launch shape
+  is pointed at reconverge, which this project ships as a component and does
+  not reimplement.
+
 ### Fixed
 
 - **`launchbound-runner` accepted a malformed `--budget-secs` in silence,

CONTRIBUTING.md

@@ -3,6 +3,23 @@
 Thanks for your interest. Issues and small PRs are welcome; large features
 are better discussed first.
 
+There are three issue forms, and which one you pick matters because a result
+here depends on more than the code — the GPU, the driver, the CUDA version,
+the compute capability, the cuda-oxide commit, the nightly, and which of the
+three backends produced the number (`docs/LIMITATIONS.md`, "Results do not
+port"). The forms ask for that provenance so a report is actionable rather
+than a starting point for questions:
+
+| form | for |
+|---|---|
+| **The gate admitted or refused the wrong configuration** | a convergence verdict that is wrong in either direction — the report that matters most, since refusing a fast configuration *is* the product |
+| **A measurement looks wrong** | a timing, a ranking or a statistic that disagrees with what you observe, including the model's estimate against silicon |
+| **An admitted candidate failed to build** | the known hole where a gate-clean configuration still fails the real compile |
+
+A finding that is wrong about the *kernel* rather than the *launch shape*
+belongs in [reconverge](https://github.com/vyncint/reconverge/issues): this
+project ships it as a component and does not reimplement it.
+
 > **These four projects share one contributor pattern** — the same commit
 > rules, the same DCO, the same AI policy, the same CI and release shape:
 > [termlens](https://github.com/vyncint/termlens),