Add test seams at I/O boundaries to raise coverage to ~90%

[drogers0]

Why

Coverage was low not because of too few tests, but because the code doing the real work had no test seam: the network, the git/gh subprocess, and os.Exit were hardwired, and the program spine sat in one large main(). Those parts were structurally unreachable from tests.

Before → after (pooled statement coverage):

Package	Before	After
. (main)	24.1%	~89%
internal/upload	14.3%	~91%
internal/repo	0.0%	~93%
internal/session	81.5%	~89%
internal/cookies	75.0%	~92%
pooled total	—	90.1%

What

Strictly behavior-preserving seams (no user-facing change — same flags, output, exit codes, URLs, and headers):

• main: extract run(args, stdout, stderr, deps) int from main(); inject the I/O boundaries via a deps struct so the full CLI spine is tested without network/subprocess/exit. The upload client is built once per run (shared cookie jar), matching prior behavior. Dead handleExtractToken/handleCheckToken removed.
• upload: introduce *Client{http, baseURL} with methods so requestPolicy/finalizeUpload/getUploadToken run against an httptest server instead of a hardcoded github.com URL.
• repo: inject a command runner so remote-URL and ID parsing are testable without a real git repo or authenticated gh.
• cookies: split pure chooseSession + mapKookyCookies out of the kooky browser read (now the only untested unit).
• session: add a profileURL var seam so CheckValidity is exercised end-to-end.
• docs: architecture.md updated for the changed signatures.

Accepted coverage gaps

internal/httputil (a single const), cookies.readRawCookies (live browser read via kooky), a few unreachable defensive branches, and thin production wiring (main, productionDeps, resolveSessionCookie).

Validation

gofmt -l . clean · go vet ./... clean · go test -race ./... pass · pooled coverage 90.1%.