api: GET /banner/summary — attempts/actor/succeeded/queue counts + pressure strip (Part 1 of #55)

[gterdem]

Summary

Part 1 of #55 — backend banner feed; frontend follows on this branch.

Adds a new additive read endpoint (GET /banner/summary, ADR-0029) feeding the dashboard triage banner:

• attempt_count / actor_count — over the ADR-0070 W_STATE (24h) window, computed from firewatch_core.attempts (the D1 attempt predicate).
• succeeded_count — the correctness crux: the success set is Tier-1 verdicts UNION actors carrying a critical-severity qualifying detection — never Tier-1 alone (ADR-0070 D3 tier-attribution correction, 2026-07-16).
• queue_size (K) — actors carrying a Tier-1 or Tier-2 escalation verdict.
• top_pressure — bounded (≤ 5) (source_ip, attempt_count, span_minutes) rows, ranked by peak decayed intensity (λ̂ is used only as an internal ranking key — never itself serialized, per ADR-0035 "engine integers only").

All counts are read-only derivations over the existing detect()/decide() verdicts and the shared attempts.py module — this PR does not touch escalation/decider.py, escalation/qualify.py, or scoring/pipeline logic. routes/banner.py mirrors pipeline.analyze_ip's exact window slicing (W_STATE/W_CAMPAIGN) so this endpoint's counts are always consistent with the per-actor ThreatScore.escalation tier already shown in the triage-banner chips — the banner can never count differently than the engine.

Why the succeeded-set union (not Tier-1 alone)

Per the issue's 2026-07-16 correction: Tier 1 requires a literal ALLOW event, and no host-auth source (syslog/linux_auth) ever emits ALLOW. So a brute_force_then_login compromise on a pure SSH box is a Tier-2 verdict (its loudness is the rule's severity="critical" registration, not the tier number). Binding "succeeded" to tier == 1 would read "0 succeeded" while the compromise rule is firing — the worst possible false calm. The success set is:

succeeded := tier == 1  OR  any(detection.severity == "critical" for detection in detections)

• Traffic-source actor (ALLOW + detection) → Tier 1 → first arm.
• Mixed-telemetry actor (e.g. CEF firewall act=permitted ALLOW + SSH brute force, same IP) → also Tier 1 via the per-actor gate → first arm.
• Pure host-auth actor firing brute_force_then_login → Tier 2, Tier 1 structurally unreachable → second arm only (the must-not regression pin).
• Ordinary failed-login-only actor (no success, no critical detection) → neither arm → correctly NOT succeeded.

New surface

• firewatch_api.banner_assembler — pure aggregation module (ActorAttemptStats, compute_actor_attempt_stats, assemble_banner_attempt_summary). No I/O; unit-testable in isolation.
• firewatch_api.routes.banner — GET /banner/summary, registered in app.py.
• firewatch_api.schemas.BannerAttemptSummary / PressureEntry — new additive response shapes.
• frontend/vite.config.ts — /banner added to the dev-proxy allowlist (new route prefix).

No existing route or response shape changes shape (ADR-0029 additive rule). tests/golden/fixtures/expected_scores.json stays byte-identical (sha256 fe4787643955c920e934e3789c79f741cd8c8cde6b2adbc6540b66ff3743f31f).

Structure self-check

• banner_assembler.py: 242 lines, one concern (pure aggregation).
• routes/banner.py: 143 lines, one concern (I/O boundary + window slicing).
• Test file: 501 lines (single-issue test suite, consistent with existing test_issue_650_escalation_policy_route.py at 600 lines).

Test plan

• New tests (21) in packages/firewatch-api/tests/test_issue_55_banner_attempts_summary.py:
  • Success-set union: host-auth must-not pin, traffic-source Tier-1, mixed-telemetry Tier-1-via-unrelated-ALLOW, ordinary non-qualifying actor, qualified-but-non-critical Tier-2 (negative case).
  • queue_size: Tier-1/2 count, observed (tier=None) excluded, Tier-3 (blocked_persistent) excluded.
  • attempt_count/actor_count: matches a direct is_attempt() tally; zero-attempt actors excluded from actor_count but not from succeeded_count/queue_size; caller-supplied windowing respected.
  • top_pressure: bounded to 5, ranked by peak intensity descending, engine-integers-only fields, span_minutes derivation.
  • Route-level: empty store → all-zero summary (calm state unchanged), no store → 503, end-to-end host-auth regression pin, response shape.
• bash scripts/gates-backend.sh — green (see report below).
• Full packages/firewatch-api/tests (885 passed) and targeted new suite (21 passed) run in isolation beforehand.
• Frontend TriageBanner.tsx consumption of this endpoint — follow-on commit on this branch (ui-dev).

🤖 Generated with Claude Code