Bound Azure Pipelines selector history and publish cost

[Evangelink]

Summary

The Azure Pipelines cache backend carries forward every selector ever observed for a weak fingerprint and republishes the complete set on each new selector.

Impact

For a recurring weak fingerprint whose path set or strong selector changes over time, each miss fetches the prior selector manifest, downloads every unique historical path set, writes temporary files, and republishes the full set. Publish work grows linearly with selector history rather than remaining bounded. Azure retention and selector deduplication limit some scenarios, but a long-lived cache universe can still accumulate substantial sequential remote I/O.

Evidence

src/AzurePipelines/PipelineCachingCacheClient.cs calls GetSelectors(...).ToHashSetAsync, adds the current selector, then iterates all selectors and republishes them. The implementation already contains TODO: limit the number of selectors we store. Timestamped selector keys carry the newest set across builds.

Suggested fix

Introduce a bounded retention policy for selectors per weak fingerprint, preserving the most useful/recent entries. Avoid downloading historical path-set content that will not be retained. Add instrumentation or tests demonstrating bounded manifest size and remote requests.

Acceptance criteria

• Selector count and manifest size remain bounded under repeated selector churn.
• Existing selectors still provide expected cache hits within the retention policy.
• Publishing a new selector performs bounded remote work.
• Tests cover churn, deduplication, and eviction behavior.

[dfederm]

This behavior is intentional. Weak fingerprints are expected to be strong enough to not have an excessive number of selectors.

Do you have evidence that the selector count is too high for your scenarios? Can you share the data, and what specifically is churning that isn't being properly predicted?