fix: track Ponder fallback routing as a flag, not a URL comparison

[Danswar]

Summary

• errorLink decides whether a failed Ponder request already went to the fallback by comparing operation.getContext().targetUrl against CONFIG.indexerFallback as strings.
• This repo's own deploy config sets CONFIG_INDEXER_FALLBACK_URL to the same value as CONFIG_INDEXER_URL (the way to keep the var populated without enabling cross-environment failover). With primary and fallback string-identical, targetUrl === CONFIG.indexerFallback is true on every request regardless of which URL it actually went to, so the warn + retry branch (lines 47-52) can never execute — every network error still falls through to logger.error.
• Confirmed live in Grafana/Loki: the last 30 ApiApolloConfig network-error lines in prod were all error, none warn, despite this fallback mechanism (fix: reclassify Apollo network-error log severity #117 / fix: attribute network errors to the URL the request was sent to #121) having been deployed since 2026-07-20.
• Fix: routingLink now stamps usedFallback: isFallbackActive() at send time, and errorLink reads that flag instead of re-deriving it from a URL comparison. This correctly distinguishes primary vs. fallback attempts even when they resolve to the same URL string.

Test plan

• Reviewed against the existing retry path: on retry, routingLink re-stamps context from the (now fallback-active) state, so a second failure against the fallback correctly falls through to error — behavior unchanged from the intended design, just no longer defeated by same-URL configs.
• After merge: watch Loki for ApiApolloConfig network-error lines to move from error to warn on the next transient Ponder blip.

(Supersedes #123, closed — that PR's follow-up commit message and description used internal-sounding shorthand that doesn't belong in a public repo.)

[Danswar]

3 review passes to 0 findings. Correctness (retry lifecycle, concurrency, the string-comparison edge case this PR fixes) verified clean on the first pass. Two follow-up passes on wording caught and fixed public-repo hygiene issues — internal shorthand terminology that had leaked into a code comment, a commit message, and the PR description — replaced with generic phrasing. Final state: no outstanding issues, CI green.