fix: retain tool catalog on transient Uplink manifest fetch failure

[Staticsubh]

Summary

Closes #761.

When operations.source: uplink is configured, a transient Uplink persisted-query manifest fetch failure (network timeout, DNS failure, HTTP 5xx, retryable retry_later response) previously caused the server to replace its active tool catalog with an empty list. The server continued reporting /health UP, sent tools/list_changed to connected clients, and those clients received {"tools": []} with no error signal.

This PR fixes the bug by mirroring the existing CollectionError keep-last-good pattern for manifest fetch failures.

Root Cause

manifest_poller.rs:44–47 converted any Err from the Uplink stream into Event::UpdateManifest(vec![]) — treating a transient infrastructure failure as an authoritative "empty collection" signal:

// Before (bug): error → silently wipes catalog
Err(e) => {
    tracing::error!("error from manifest stream: {}", e);
    Event::UpdateManifest(vec![])
}

Changes

• apollo-mcp-registry — Added ManifestError(BoxError) variant to the manifest event enum. manifest_poller.rs now emits ManifestError(e) on Err instead of UpdateManifest(vec![]).
• apollo-mcp-server — Added ManifestError(BoxError) to ServerEvent and OperationError. operation_source.rs forwards ManifestError through the event stream. State machine handles ManifestError while Running by retaining the existing catalog and logging the error (exactly as CollectionError is handled); treats it as fatal during startup before the first successful load.
• Tests — Two new regression tests in states.rs: one confirming the server stays alive with its catalog intact on ManifestError while Running; one confirming it is fatal during startup.

Behavior

Uplink response	Before	After
HTTP 200, non-empty manifest	Apply catalog ✅	Apply catalog ✅
HTTP 200, empty manifest	Clear catalog ✅	Clear catalog ✅
Network error / timeout	Silent empty catalog ❌	Retain catalog, log error ✅
HTTP 5xx / retry_later	Silent empty catalog ❌	Retain catalog, log error ✅

Test Plan

• cargo test -p apollo-mcp-server manifest_error — both new regression tests pass
• cargo test -p apollo-mcp-server collection_error — existing tests still pass
• cargo clippy --all-targets -- --deny warnings — no new warnings
• Manual: start server with operations.source: uplink, simulate Uplink unreachable, confirm tools/list returns previous tools and /health logs a structured error

🤖 Generated with Claude Code

[apollo-librarian]

✅ AI Style Review — No Changes Detected

No MDX files were changed in this pull request.

Review Log: View detailed log

This review is AI-generated. Please use common sense when accepting these suggestions, as they may not always be accurate or appropriate for your specific context.

[DaleSeo]

Thanks for the fix, @Staticsubh. This looks great! I have a few comments.