refactor(core)!: panic-free builder validation — build() collects all configuration errors (034 Phase 2, #133)

[lxsaah]

Closes #133. Part of design 034 Phase 2 (review doc §3.4); second of three stacked PRs (#130 → #133 → #134). Builds on #137.

What

One failure model instead of two: builder methods never panic on user mistakes; build() validates everything and returns all errors at once.

• New public ConfigError { record_key, url, message } + DbError::InvalidConfiguration { errors: Vec<ConfigError> } (code 0x4003, included in is_configuration_error).
• TypedRecord::set_producer / set_transform / add_inbound_connector: mutual-exclusion violations and duplicate producer/transform are recorded and the conflicting registration is skipped (has_producer() stays false after a conflicting .source()).
• OutboundConnectorBuilder::finish() / InboundConnectorBuilder::finish(): invalid URL, missing serializer/deserializer, unregistered scheme, transform/source conflicts → recorded with record key + URL; the link is not registered; the registrar is returned as usual so chains keep compiling.
• Spawn-time factory panics become build()-time checks: a record with an outbound/inbound link and no buffer now fails build() with key + link URL (previously: panic inside a factory closure during the connector phase, far from the user's mistake). Consequence: .buffer() after .link_to()/.link_from() is now legal — order-independent (covered by a new test).
• configure() re-registering a key with a different type: recorded, user closure skipped (was assert!).
• Duplicate keys and DependencyGraph findings (cycles, unregistered transform inputs) fold into the same collected report. Note for reviewers: callers matching DuplicateRecordKey/CyclicDependency from build() must switch to InvalidConfiguration — flagged per the issue's directive to collect all errors.
• Remaining panic!/expects on the builder path are internal invariants only, worded "this is a bug in aimdb-core".
• AnyRecord gains has_buffer() and drain_config_errors() (breaking for external implementors).

Acceptance criteria (from #133)

• Grep of panic!/assert! in typed_api.rs/typed_record.rs/builder.rs shows only internal-invariant messages
• A builder with three distinct mistakes reports all three from one build() call (build_reports_all_configuration_mistakes_at_once)
• Every error carries the record key and (where applicable) the connector URL
• Error construction is single-shape per refactor(core): unify DbError on alloc::String — remove std/no_std dual-field variants #129
• make check passes; examples unchanged

Tests

The 6 #[should_panic] tests are converted to recorded-error assertions; new tests cover unregistered-scheme recording, the three-mistakes acceptance criterion, buffer-after-link order independence, linked-record-without-buffer at build(), and InvalidConfiguration's Display/error-code. The tokio adapter's test_duplicate_key_error now asserts the collected error instead of a configure-time panic.

🤖 Generated with Claude Code