Milestone 3.5.C.6 — gRPC Service Skeleton and Degradation Contract

Status: Planned
Goal: Track C — Context Assembly Engine
Phase: 3.5 — Extraction & Context Assembly
Estimated effort: 3 days
Track: Track C — Context Assembly Engine
Depends on: Milestone 3.5.C.5 (context formatter)

Why This Milestone Exists

Wraps Track C's budget/retrieval/scorer/packer/formatter internals into the actual ContextAssemblyService gRPC server, and locks down exactly what "graceful degradation" means operationally.

Degradation Ladder

Level	Trigger	Behavior	Response signal
L0 — Full	All retrieval sources respond within budget	Directive + hot + cold memories, full scoring/packing	`context_assembly_fallback=false`
L1 — Cold-search degraded	Cold pgvector/HNSW search exceeds its sub-deadline	Hot cache + directive only, cold results dropped	`AssemblyMetrics.candidates_evaluated` reflects hot-only count
L2 — Memory degraded	Both hot and cold sources fail/timeout	Directive-only context	`was_truncated=false`, `memories_included=0`
L3 — Full degraded	gRPC call itself times out (45ms) or connection refused	Proxy uses original, unmodified messages	`X-IBEX-Context-Fallback: true`

Orientation (indicative)

Named paths, package layouts, libraries, schemas, env vars, and commands anywhere on this page are rough sketches for orientation — inspiration and a baseline, not a required change list.

During implementation, expect to:

open the live tree and follow existing patterns before inventing new ones
research current constraints (latency, tenancy, deploy shape, libraries) more deeply than this page can
advance the design beyond the sketch where measurement or code reality says so
land work in different filenames, merged packages, deferred docs, or new surfaces when the situation calls for it

Prefer outcomes over matching any particular file tree or command sequence.

Areas that may be involved (situational — not a checklist):

Context assembly service
Tokenizer registry / counting

Success signals

Outcome-oriented signals that the milestone is in good shape. Exact filenames, package layouts, and commands may differ from any sketches above.

Each degradation level has a dedicated unit test with a fault-injected dependency
AssemblyMetrics populated accurately at every level (not just L0)
GRPC server enforces its own internal 40ms retrieval deadline independently of the proxy's 45ms client-side timeout
Load test: sustained 500 RPS against AssembleContext with p99 < 50ms on a 100K-memory corpus
Repo guards / CI checks still pass

Prerequisites

Milestones 3.5.C.1–3.5.C.5