[llima]

Summary

• Version bumped to 0.34.0 (a minor, spec 38 Ruling 6). Merging publishes craftar@0.34.0 to npm after approval in the npm environment.
• craftar init --new-profile <name> [--like <profile>] creates the profile first — one new file, profiles/<name>/profile.yaml, in the Forge's working tree, holding name, recipes and targets only — then starts the workspace on it. Nothing is committed, no branch is made and nothing is pushed (Ruling 1). With --new-profile, --add-recipe, --remove-recipe, --replace and --targets shape the profile, and craftar.yaml gets no recipes and no targets.
• --like copies the template's written recipes and targets only (Ruling 2). In a Forge at schema 3 or above, recipes that only the template's integrations block selects are left behind with one note (Ruling 3). A recipe named for another profile is refused (P10, Ruling 4).
• init becomes a gated Forge writer of one path. The gate (P5–P8) refuses a cache tree, a directory that is not a git repository with a commit, a name or profiles/<name> that exists, a path git ignores, and a symbolic link on the way. The profile directory is journalled, created with a non-recursive mkdir, and the file is written wx. The Forge is then read back from disk and must plan the same bytes before craftar.yaml is written. A late failure prints git -C <forge> clean -fd -- profiles/<name>.
• In a terminal, the profile list offers 0) a new profile, and a Forge directory with no profile goes straight to the new profile's questions (Ruling 5).
• No emitted byte, lock, registry entry or --json shape changes. No Upgrading block.

Built as spec 38 recommends, not yet ruled (§13)

• §13.1 items 2, 3, 4, 5, 9 and 10.
• §13.2 items 11–21.
• Item 4 is the second plan from disk: one more load-by-source, on purpose. The workspace rule node-cli.md and its reviewer agent gain that sentence after this PR, together with init as a gated writer.

Departures from spec 38's text (recorded at the spec refresh)

• The writes of §5.3 live in a core module (src/core/new-profile-write.ts) with injected seams (WriteHooks), so the race and the second proof are tested in-process. There is no environment back door.
• P12 is the resolver's own refusal; a missing ingredient stays a warning.
• P5 fails closed on any error except a missing forges/.
• Names YAML would misread (123, null, true) are quoted in the file, and the file is read back through ProfileSchema before any write.
• A local recipes or targets is said, not asked, and stays a workspace layer.
• Every refusal in the interactive recipes step is re-asked. The one exception is a template that brings an owned recipe while craftar.local.yaml sets recipes: that ends on P10, as the flags mode does.

Test plan

• npm run typecheck — exit 0 (0f7e031, after the bump and a rebase onto 6eeaeae — the 25 commits identical by range-diff)
• npm run build — exit 0; craftar --version prints 0.34.0
• npm test (vitest --maxWorkers=6, without test/ci.test.ts) on 0f7e031 — 136 files passed, 1 skipped; 3366 tests passed, 10 skipped. Oracle skipped (no fixture); no emitter, plan or status byte path changed.
• Pins written first (test/init-new-profile-pins.test.ts): init's both-flags stdout byte for byte, and today's re-ask on the answer 0.
• Every new test was seen failing first against unchanged src/.
• The late failures at the CLI were produced with read-only directories (Linux only, skipIf(win32)).
• CI green on Linux and Windows.
• Coded, not tested: craftar.yaml losing its wx race to N1 after the profile was written. There is no seam between the two writes.
• Not run by the suite, and not tested by hand: the interactive mode in a real terminal (spec 28's gap) — the 0) entry, the questions, the re-asks, the confirmation.
• By hand, not run: two real processes racing on one name; the link-to-a-file case on Windows; P8 on a case-insensitive disk.