Restructure AGENTS.md into scoped guides

Trim AGENTS.md to whole-repo, every-session facts (154 -> 110 lines) and move
subsystem guidance into docs/guides/*.md, each with paths: frontmatter and a
one-line index entry in AGENTS.md. Symlink each guide from .claude/rules/ so
Claude Code lazy-loads it by matching path; track those symlinks via a
.gitignore exception (.claude/settings.local.json stays ignored).

docs/guides/testing.md

@@ -0,0 +1,59 @@
+---
+paths:
+  - "licensing-service-startos/licensing-service/tests/**"
+---
+
+# Testing, lint & CI status
+
+## Running
+
+From `licensing-service-startos/licensing-service/`:
+
+```
+cargo test                  # all suites
+cargo test --test api       # one suite — suites: api crosscheck migrations subscriptions upgrades worker
+cargo test <name>           # single test by name pattern
+cargo clippy --all-targets  # ~6 style lints (e.g. unnecessary_map_or)
+cargo fmt --check           # reports formatting diffs (does not write)
+```
+
+## There is no CI
+
+No `.github/workflows`, no Gitea/Woodpecker/Drone config. **Nothing is enforced** —
+not `cargo test`, not `clippy`, not `fmt`. The release gate is `publish.sh`.
+
+## Formatting is not clean and not enforced
+
+The tree has never been run through `cargo fmt` (56 daemon files differ) or
+`prettier` (37 wrapper files differ). `cargo check` and `tsc --noEmit` pass clean.
+If you ever format, do it as a **standalone commit** so the mechanical churn
+doesn't bury logic changes.
+
+## Runtime-prepared SQL → tests are the only safety net
+
+`db/repo.rs` queries are `sqlx::query(&format!(...))`, prepared at runtime, so bad
+or ambiguous columns 500 only when executed. A real prod regression (every paid
+purchase returning 500 on `:52`) shipped this way because the new merchant-profile
+resolution path had no passing test. When adding/altering a repo query, add a test
+that actually executes it. See [payments](payments.md).
+
+## Known-failing tests (3 in tests/api.rs)
+
+Test-debt from the `:52` payments transition; the backend is sound, these are a
+test-strategy decision, not bugs:
+
+- `paid_purchase_creates_invoice_via_provider`, `paid_purchase_in_usd_records_listed_currency_and_rate`
+  — fail with a legitimate 400 ("no payment providers connected"); the fixture
+  seeds no provider. Reaching 200 needs the mock wired through
+  `resolve_provider_for_profile_rail` (`build_provider` only makes real clients;
+  `#[cfg(test)]` does not apply to integration tests against the lib).
+- `payment_provider_preference_round_trip` — inserts into the dropped
+  `btcpay_config`/`zaprite_config`; superseded by
+  `merchant_profile_provider_resolution_queries_round_trip` (delete or rewrite).
+
+The other 43 api tests + all other suites pass.
+
+## Cross-language wire-format tests
+
+The LIC1 format is verified across SDKs separately — see
+[crypto-wire-format](crypto-wire-format.md).