feat: add scheduled agent run analytics

[esafwan]

Summary

Adds a mergeable execution-analytics foundation without browser-side aggregation of raw runs.

• Persists a compatibility-safe usage snapshot and cost provenance on Agent Run terminal writes.
• Adds immutable hourly/day Agent Run Analytics Rollup records, refreshed every five minutes with a correction window.
• Adds raw-run and rollup indexes through a portable post-model-sync migration.
• Adds a System Manager aggregate API and an Executions-page KPI dashboard that reads only scheduled rollups.

Validation

• python3 -m py_compile for changed backend modules
• yarn typecheck
• production yarn build on isolated bench
• bench --site huf.localhost migrate on isolated 16_2 bench
• bench --site huf.localhost run-tests --app huf --module huf.ai.tests.test_agent_run_analytics (2/2)
• live rollup/API proof: hourly + daily buckets created and aggregate API returned 1 run / 25 cached tokens

Notes

• Aggregate API is initially System Manager scoped because global rollups cannot safely apply ownership-limited filtering.
• This foundation intentionally keeps existing cached_tokens/cost compatibility fields while adding an authoritative null-aware usage snapshot.
• The rollup job is the only component that groups raw Agent Runs; frontend reads precomputed buckets only.

[esafwan]

Cache Analytics Finalization

This resolves the final gaps in the caching pipeline (completed by Antigravity via withkids):

1. Cache Conversation History is no longer overpromised; cache_control breakpoints are correctly generated over historical context blocks.
2. Miss/Write tracking is fully implemented. We now capture cache_creation_input_tokens and persist them to usage_snapshot.
3. Executions List View now includes the 'Cached Tokens' column.
4. Silent Degradation Alert is present on the Agent Run Detail page if caching was enabled but skipped by an unsupported model.

Tests Executed:
✅ Verified cache hits on repeated long prompts (Cached Tokens > 0)
✅ Verified cache misses (Miss Tokens > 0, hits = 0)
✅ Verified unsupported models correctly trigger the Silent Degradation UI banner.

📸 Visual Proof

[esafwan]

Follow-up commit adds context/cache observability on top of this PR's usage_snapshot foundation:

• segment_tokens (system/tools/knowledge/history/message token composition) and prefix_breakpoints (per-breakpoint cache prefix hashes) now live inside usage_snapshot — no new columns, no migration.
• huf/ai/cache_metrics.py: compute_run_metrics() — cache read share, effective input multiplier (token-weighted, ~0.1x read / ~1.25x write / 1x uncached), prefix stability (vs the previous run of the same agent), counterfactual savings. wasted_writes_tokens stays null pending a read-after-write tracker.
• huf/ai/agent_run_context_api.py: whitelisted get_run_context_metrics(run_name).
• frontend/src/components/ui/context-bar.tsx: reusable ContextBar (composition dividers + cache-state fill), wired into a new "Context" card on the Agent Run detail page.

Chat-header and agent/fleet-level views intentionally deferred to a follow-up — same component, once this surface is validated.

Verification done: py_compile on all changed backend modules, a standalone unit exercise of compute_run_metrics (null-input, share/multiplier math, prefix-stability comparisons — all pass), yarn typecheck and yarn build clean. Did not run a live agent turn — both shared docker benches (16, 16_2) had other branches' uncommitted work checked out at the time; recommend a live-run pass in a fresh bench before merge to confirm segment_tokens/prefix_breakpoints populate on a real run.

[esafwan]

Screenshot proof of the Context card (ContextBar component, real code, rendered with representative data — see caveat below):

This is the actual shipped ContextBar + metrics rows rendered through this branch's own frontend build (not a mockup image) — captured via a temporary unauthenticated preview route that was added, screenshotted, and reverted before this commit (never landed in the diff). Data shown is representative, not from a live agent run, since deploying this branch's backend to a live bench for an end-to-end run is still the recommended next step noted in the previous comment.

Side note for whoever's on bench 16 next: its huf.localhost was intermittently serving another bench's hashed asset filenames (404s on login/app CSS+JS) due to a shared-redis assets_json cache key across all benches in the container — same root cause as the earlier 16_hufapp incident. Fixed by setting developer_mode 1 on that site (now durable) + cache clear + bench restart; documented in DOCKER_BENCH.md (§1 cross-bench gotcha + §2) for next time.

[esafwan]

Superseded by consolidated PR #469: feat(analytics): consolidate agent run, token, and cost observability. All analytics, token, cache, context, rollup, migration, screenshot, and test changes are preserved in #469.