Story-2081 Walkthrough

Change: This is S-4 of the agent prompt-caching build (Epic-2050, RFD docs/notes/rfd-agent-caching-strategy.md §9 / D-8; sibling of pramaan-functions rfd-context-engine.md §14.3 S-1). The send-twice probe harness shipped earlier (Spike-2055, PR #191) but ran with no key — the matrix was PENDING LIVE RUN. This story is the live run itself: execute the harness against OpenRouter with a real ZDR-pinned dev key, turn every _pending_ row into observed wire behavior, and feed the verdict into paid_write_providers. No code change — a measurement story: the harness is unchanged, only the committed matrix and two config annotations move.

What ships

File	Role
`docs/notes/zdr-caching-probe-matrix.md`	The `PENDING` tables replaced with the live 2026-07-07 matrix: per-model verdict, `c1`/`c2` cached tokens, hit rate, provider stickiness, plus the three §9 assertions resolved and the hypothesis-vs-reality read.
`agents/framer/llm.yaml`, `agents/mira/llm.yaml`	`paid_write_providers` annotated with the live provenance — `anthropic` empirically confirmed paid-write, `alibaba` retained defensively. Value unchanged (`[anthropic, alibaba]`); the probe validated the existing list.
`docs/notes/zdr_probe.py`	Unchanged — the Spike-2055 harness ran verbatim. Listed only to name what was executed.

The run (one command each, harness-only evidence)

export OPENROUTER_API_KEY=<dev key>
python docs/notes/zdr_probe.py           # ZDR-eligibility ping
python docs/notes/zdr_probe.py --cache   # send-twice caching matrix (nonce=60477-1783412227)

Every call carried the hard pin {"zdr": true, "data_collection": "deny"} and a per-model session_id. The nonce is baked fresh into the ≥4,096-tok synthetic prefix, so call 1 is a cold write (c1_cache≈0) and call 2 the read (c2_cache>0 on a HIT) — cold-write-then-read within one run, no reliance on a prior warm cache. Synthetic fixtures only (RFD §9 cl.5).

Observed matrix (live)

Model	Expected	Verdict	c2 cache tok	hit	provider (1→2)	sticky
`openai/gpt-5.4`	auto	CACHE_HIT	3840	0.927	Azure	✓
`anthropic/claude-sonnet-5`	explicit	CACHE_HIT	7046	0.997	Amazon Bedrock	✓
`google/gemini-3.1-flash-lite`	implicit	MISS	0	0.0	Google	✓
`deepseek/deepseek-v4-pro`	unknown	MISS	0	0.0	SiliconFlow	✓
`minimax/minimax-m3`	unknown	CACHE_HIT	4224	0.999	DeepInfra	✓
`qwen/qwen3.6-35b-a3b`	unknown	MISS	0	0.0	AtlasCloud	✓
`bytedance-seed/seed-2.0-lite`	unknown	MISS	0	0.0	Seed	✓
`z-ai/glm-5.2`	unknown	MISS	0	0.0	DeepInfra	✓
`xiaomi/mimo-v2.5`	unknown	CACHE_HIT	4224	0.994	DigitalOcean	✓
`arcee-ai/trinity-mini`	unknown	MISS	0	0.0	Clarifai	✓

What the numbers say

Frontier caches confirmed under the ZDR pin. Azure-OpenAI auto-cache (gpt-5.4) and Bedrock-Anthropic explicit-breakpoint cache (sonnet-5) both HIT — RFD §9 assumptions 1 & 4 move from assumed to observed. OQ-2 answered: gpt-5.4 is servable + auto-caching live, so it stays the converse-primary candidate.
"Aggregator ⇒ no cache" is false (assumption 2 partially refuted): DeepInfra (minimax-m3) and DigitalOcean (mimo-v2.5) HIT. And it is per host×model, not per host — z-ai/glm-5.2 also on DeepInfra MISSed. The lesson: verify each route, don't assume. Exactly why this is a probe, not a spec table.
Gemini implicit MISSed back-to-back — an implicit cache populates after the first response materializes; a near-instant second call can precede the write. Not "never caches"; re-test with a delay before leaning on it.
Sticky routing held on all 10 rows and every served provider is ZDR-eligible by construction (the zdr flag rides every request). RFD §9 assumption 3 confirmed; no ZDR escape via a sticky fallback.
cache_discount is absent from the wire — OpenRouter returned no such field on any route even with usage:{include:true}. Recorded as absent (the assertion is "recorded", not "non-null"); S-10 must derive realized-$ from provider_details["cost"] deltas, per RFD §10.

Verdict → `paid_write_providers`

Rule (RFD §6.1): paid-write iff the provider serves an explicit-breakpoint cache. Only anthropic (sonnet-5/Bedrock) qualified live → CONFIRMED. alibaba/qwen served no cache on its AtlasCloud route so no write was billed, but Qwen's upstream cache is documented explicit, so it is kept defensively — cheap to keep, unsafe to drop. The auto-cache HITs (openai/minimax/xiaomi) are free-write and correctly excluded. The live run validates the existing [anthropic, alibaba] in both llm.yamls; no value change, only annotated provenance. S-9's ambient guard test enforces that ambient seats never emit markers on this set.

Acceptance criteria → evidence

AC	Proof
matrix committed	`docs/notes/zdr-caching-probe-matrix.md` — every `_pending_` row replaced with the live 2026-07-07 send-twice result.
`cache_read_tokens > 0` harness-logged	4 rows with raw call-2 `cached_tokens`: gpt-5.4 (3840), sonnet-5 (7046), minimax-m3 (4224), mimo-v2.5 (4224). The other 6 honestly read 0.
provider ∈ ZDR set harness-logged	All 10 served providers returned under the `zdr` pin; companion ping shows the filter fails closed on non-ZDR ids.
`cache_discount` recorded	Recorded as absent across all rows — the wire returned no discount field; realized-$ deferred to S-10 via billed cost.

Measurement story, no runtime surface: the harness is unchanged, no agent code moved, no CodeZip rebuild or AgentCore redeploy. The one design consumer (paid_write_providers) was validated, not changed. Both RFDs' wiring (OpenRouterModelSettings + CachePoint()) is unaffected — the probe informs which seats profit, never whether the strategy is built (issue AC-3).