refactor(env): rename E2A_BASE_URL to E2A_API_URL; split it from the CLI's E2A_URL

[jiashuoz]

Why

The client-side base-URL env vars named one concept badly and two concepts identically, so following our own docs silently misrouted traffic.

E2A_URL and the SDKs' E2A_BASE_URL are not the same thing:

• The CLI's api_url is the deployment root. login opens a browser against it (cli/src/commands/login.ts:310) and points users at /get-started — a Next.js route (web/src/app/(app)/get-started) the Go backend does not serve. https://e2a.dev works because the web front also proxies /v1 (web/Caddyfile:35).
• The SDKs and the MCP server only ever call /v1, and want the API host alone.

Two names for "the base URL", with two different defaults, hid that distinction. The concrete failures:

Docs taught a var the SDKs don't read	README.md and docs/deployment.md teach E2A_URL; the SDKs read only E2A_BASE_URL. A self-hoster who followed the docs got their SDK silently pointed at production api.e2a.dev.
The CLI alias broke e2a login	The CLI accepted E2A_BASE_URL as an alias for its deployment root, so exporting E2A_BASE_URL=https://api.e2a.dev to configure an SDK silently repointed the CLI at the API host.
tether forced the CLI onto the API host	tether.env.example taught E2A_BASE_URL ("self-host: your API base") and lib.sh:51 defaulted it to api.e2a.dev — overriding the CLI's own e2a.dev default through that alias.

What changed

• SDKs (TS + Python) read E2A_API_URL, matching the name the server already uses for the same concept (internal/config/config.go:344 — its externally visible API base). E2A_BASE_URL still works, with a deprecation warning (console.warn / DeprecationWarning).
• MCP moves to E2A_API_URL (E2A_URL + E2A_BASE_URL deprecated) and now defaults to the API host — it is a pure API client.
• CLI keeps E2A_URL alone and no longer reads E2A_BASE_URL. Dropping the alias silently would strand tether self-hosters on production, so it warns — gated to the actually-dangerous case (legacy var set, E2A_URL unset, root still at default), so a legitimate SDK user who also configured their CLI gets no noise.
• tether migrates to E2A_URL, carries a legacy ~/.e2a-tether.env over with a warning (rather than reverting to prod), unsets the stale name so the CLI doesn't double-warn, and tracks the CLI's default instead of forcing api.e2a.dev.
• Docs: dropped the E2A_AGENT_EMAIL fallback claim from both SDK READMEs and the root README — it was never implemented in either SDK (they raise missing_email), though it is real in the CLI and tether. Added the CLI's env-var table (it documented none despite reading five) and split E2A_URL vs E2A_API_URL in docs/deployment.md.

Also removed the MCP one-shot deprecation guard: main() calls loadConfig exactly once, so it deduped nothing and only leaked state across tests.

Compatibility

Additive. Every previously-working configuration still works; the old names warn rather than break. The one behavior change is the CLI ignoring E2A_BASE_URL — loud, not silent, and tether (its main consumer) migrates itself.

Verification

Suites: TS SDK typecheck+unit+types, MCP vitest+types, Python 323 passed / 18 skipped, CLI 120 passed, tether bash -n clean.

Driven for real, not just unit-tested:

• Both SDKs against actual dialed origins across all four resolution cases (default → E2A_API_URL → deprecated E2A_BASE_URL + warning → both set, canonical wins).
• The built CLI binary: warning fires for the self-hoster case, silent when E2A_URL is set.
• t_load_config through the legacy-migration path: carries over, warns, unsets, and defaults correctly.

⚠️ pytest inside a git worktree silently tests the main checkout via the editable install — run it with PYTHONPATH=$(pwd)/src or you are testing the wrong tree.

Follow-ups (not in this PR)

• mcp/src/config.ts is dead code (nothing in src/ imports it but a type; no bin field since npm stdio publishing was retired in GA /v1 readiness: settle API contract, regenerate SDKs, slim CLI, re-curate MCP (hosted-only) #247) and now also contradicts the new canonical name. Deleting it is a separate call.
• The split-host topology internal/config/config.go:377 describes ("web on one host, API/MCP on another") still can't be expressed by the CLI, which holds only one URL. That's the deeper fix.

🤖 Generated with Claude Code

[jiashuoz]

Both addressed in 6029c08.

[P1] Docs no longer misroute SDK users. The CLI / SDK user row and the single export E2A_URL walkthrough were the misdirection — the E2A_API_URL row I'd added lower down didn't undo the headline instruction. Split into two audiences with two setup examples, in both the docs/deployment.md table and the README table:

• CLI user → E2A_URL + e2a login
• SDK / MCP user → export E2A_API_URL=… + E2A_API_KEY

Each section now names the var the other audience's tooling ignores, e.g. "The CLI does not read E2A_API_URL … so a self-hoster who only exports E2A_API_URL leaves the CLI pointed at production."

[P2] Legacy override is no longer silently ignored. You're right that gating on resolved === DEFAULT_URL missed the stored-host case: E2A_BASE_URL used to override ~/.e2a/config.json, so stored host A + legacy env host B silently kept A. The warning now fires whenever E2A_BASE_URL is set and E2A_URL is absent, and names the host actually in use so the mismatch is visible. Driven against the built binary with a seeded non-default api_url:

e2a: E2A_BASE_URL is set but the CLI does not read it — that name configures the SDKs.
     The CLI uses E2A_URL for the deployment root and is talking to https://stored.selfhost.dev.
     Set E2A_URL to point it at a self-hosted deployment.

Regression test added with a stored non-default api_url asserting both the warning and that it names the stored host. CLI suite 121 passing.