feat(cli): harness init scaffold, BYOH upstreaming flow, and descriptor PR template

[slowdini]

Closes #138 (part of the descriptor epic #122, step 7 — make the BYOH path agent-followable).

What changed

• New subcommand eval-magic harness init <name> — scaffolds .eval-magic/harnesses/<name>.toml, a fully commented descriptor template (only label live; every optional table a commented-out, inline-explained example with VERIFY-don't-guess prompts) plus <name>-notes.md, the verification-record skeleton. The scaffold is lint-clean as written and registers the name as a baseline harness immediately. --stdout prints the template for redirection, --force overwrites, non-kebab names are rejected, and scaffolding a registered name (e.g. claude-code) prints an overlay note instead of erroring.
• Embedded template assets harnesses/template.toml / harnesses/template-notes.md, packaged with the crate (pinned by the package test). Unit drift-guards enforce that the template mentions every schema field and named-capability value, and that the fully-uncommented examples form one coherent, invariant-clean descriptor.
• docs/byoh.md: new "Upstreaming your descriptor" section (exactly which capabilities are data vs code, the four-file data-only PR shape, how to open it with the template) and a "Verify, don't guess" section; the five-minute version and workflow now lead with harness init.
• New PR template .github/PULL_REQUEST_TEMPLATE/harness-descriptor.md for data-only descriptor contributions: scope check (existing named capabilities only, no [guard]), expected diff shape, verification evidence, and the don't-guess attestation. Ordinary PRs stay template-less.
• Docs touches: progressive-enhancements.md ("Adding a new harness" + Guardrails reference the scaffold and PR template), README Harnesses/Documentation/Bundled-assets sections, and the top-level --help examples.

Before / after

Before, the five-minute BYOH flow started from a bare heredoc:

mkdir -p .eval-magic/harnesses
cat > .eval-magic/harnesses/cool-custom-harness.toml <<'TOML'
label = "cool-custom-harness"
TOML

After:

$ eval-magic harness init cool-cli
Scaffolded harness descriptor: .eval-magic/harnesses/cool-cli.toml
Scaffolded notes skeleton:     .eval-magic/harnesses/cool-cli-notes.md

✓ TOML syntax + schema
✓ user-layer restrictions ([guard] stays built-in-only)
✓ cross-field invariants
Linted .eval-magic/harnesses/cool-cli.toml: all checks passed.

Next:
  1. Fill in verified values — follow the template's comments; never guess a flag
     (record each value's source in .eval-magic/harnesses/cool-cli-notes.md).
  2. Re-lint after every edit: eval-magic harness lint .eval-magic/harnesses/cool-cli.toml
  3. Smoke eval: eval-magic run --harness cool-cli  (dispatch, then ingest + finalize)
  4. Upstreaming the proven descriptor: docs/byoh.md "Upstreaming your descriptor".

Verification

• cargo test — 430 lib + 94 cli + 69 run tests pass (10 new integration tests in tests/cli/harness_init.rs, 4 new unit drift-guards; both guards mutation-tested)
• cargo fmt --check and cargo clippy --all-targets -- -D warnings clean
• Manual: harness init cool-cli → lint → harness list (shows cool-cli project baseline), overlay note on claude-code, --stdout mode, harness init --help, top-level --help examples block

Notes for review

• CLI surface addition (harness init) with thorough --help; no schema changes; docs updated across byoh.md / progressive-enhancements.md / README.
• --probe is still only "planned" everywhere — the template and docs never print it as runnable.

🤖 Generated with Claude Code