feat: add @braidhq/source-loader-github for GitHub issues

[mroops0111]

GitHub issues are intent: a feature request or bug report is
exactly the "why we built this" content that drives braid-extract.
Today intent sources are filesystem or MCP; this issue adds a third
option that pulls issues from a GitHub repo into a filesystem
layout the existing skills can already read.

Independent package, not bundled into
@braidhq/source-loader-git. The git loader clones repos; this
loader hits the REST API. Auth, config shape, and sync semantics
share almost no code. GitLab and Bitbucket equivalents will be
separate packages by the same logic; whether a shared base layer
is worth extracting will be obvious after three providers exist.

Scope

• New package @braidhq/source-loader-github, plugin
  kind: 'github'.
• Config: { owner, repo, state?, labels?, includeComments?, includePullRequests? }. Token via ${GH_TOKEN} env
  interpolation, same pattern as source-loader-git.
• ingest: REST GET /repos/:owner/:repo/issues, paginated; each
  issue becomes <destination>/issues/<number>.md with
  frontmatter (title, state, labels, author, createdAt,
  updatedAt, url) + body + a ## Comments section.
• sync: same call with since=<lastFetchedAt>. Output is
  diff-stable: deterministic key order in frontmatter; comments
  sorted by createdAt.
• Register the plugin in composeFs.ts when GH_TOKEN is set,
  mirroring the gdrive OAuth gating.

Out of scope

• GitLab / Bitbucket loaders.
• PR reviews / inline diff comments. includePullRequests in v0
  only adds title + body + top-level comments.
• Webhook ingest. Separate issue.
• Comment-thread reconstruction; flat list is fine for v0.

Acceptance

1. Plugin unit test with stubbed fetch: ingest produces N
   markdown files for N fixture issues; frontmatter parses;
   content matches.
2. sync with a since cursor only rewrites changed issues;
   untouched files stay byte-identical between syncs.
3. Dogfood: register this repo as a github source on a local
   workspace; manually run /braid-extract on the synced issue
   directory; resulting proposals look sensible.

Depends on: #27 — needs IntentExtractionLedger API frozen.