eval: capture v0.1.5 baseline via config-emulation

Replaces the placeholder v0.1.5.json (all-zeros, _baseline_pending=true)
with measured numbers so the bench harness can compute real
baseline_delta values.

Methodology — config-emulation rather than v0.1.5 source checkout:
the runner calls TunedHybridBackend.query() with no where-filter, so
Phase 7 classifier, Phase 9 valid_to, and Phase 11 filtered_dense
levers don't fire at query time regardless of the indexer-side
payload schema. Running the current source with retrieval.reranker=off
therefore reproduces v0.1.5 retrieval semantics on the existing
collection, and a checkout dance is unnecessary.

Source measurement: ~/.supamem/eval/v0.3.0a3-full-20260504T103341Z.json
(470 records, heuristic judge, full LongMemEval_S). The capture_method
and capture_notes fields document this in the baseline JSON itself so
future delta claims can cite the methodology.

Headline finding (now visible in baseline_delta):
  current default v0.3.0a3 stack (rerank-on) vs this baseline:
  tokens_per_correct_answer 1374.59 -> 1510.34 (+9.9% regression)

The milestone gate (-30% reduction) is currently failing by a wide
margin; Phase 13 is formally blocked per the ROADMAP gating rule.

Author: dzmitrys-dev

src/supamem/eval/baselines/v0.1.5.json

@@ -1,19 +1,85 @@
 {
   "version": "v0.1.5",
-  "captured_at": "2026-05-03T00:00:00Z",
+  "captured_at": "2026-05-04T16:51:15Z",
+  "capture_method": "config-emulation",
+  "capture_notes": "Captured 2026-05-04 via config-emulation rather than v0.1.5 source checkout. Methodology: current source (v0.3.0a3) with retrieval.reranker=off, no classifier where-filter, no temporal filter, default tuned_hybrid backend. The existing supamem-supamem collection contains v0.3.0a3 payload metadata, but classifier and valid_to fields are unused at query time when the harness calls backend.query() with no where filter. This emulates pre-Phase-8 retrieval semantics. Source eval JSON: ~/.supamem/eval/v0.3.0a3-full-20260504T103341Z.json (470 records, heuristic judge). Phase 10 success criterion gated on -30% tokens_per_correct_answer reduction vs this baseline.",
   "suite": "longmemeval_s",
-  "judge": {"kind": "heuristic", "model": "n/a"},
+  "judge": {
+    "kind": "heuristic",
+    "model": "n/a"
+  },
+  "dataset": {
+    "name": "longmemeval_s",
+    "revision": "98d7416c24c778c2fee6e6f3006e7a073259d48f",
+    "n": 470,
+    "subset_ids": []
+  },
   "scores": {
-    "recall_at_5": 0.0,
+    "recall_at_5": 0.21693683703050703,
     "context_precision": null,
     "context_recall": null,
     "answer_relevance": null,
-    "tokens_per_correct_answer": 0.0,
-    "context_compression_ratio": 0.0,
-    "input_tokens_p50": 0.0,
-    "input_tokens_p95": 0.0,
-    "write_cost": 0.0
+    "tokens_per_correct_answer": 1374.5881560283688,
+    "context_compression_ratio": 384.6544559994397,
+    "input_tokens_p50": 742.0,
+    "input_tokens_p95": 920.0,
+    "write_cost": 746.9063829787234
   },
-  "_baseline_pending": true,
-  "notes": "Pending real capture against the v0.1.5 release tag with heuristic judge. Plan 10-06 (release) replaces these placeholders with measured numbers before tagging. The runner's load_baseline tolerates this stub via the _baseline_pending flag and emits an err_console hint."
+  "by_axis": {
+    "single_session_user": {
+      "recall_at_5": 0.1008008658008658,
+      "context_precision": 0.0,
+      "context_recall": 0.0,
+      "answer_relevance": 0.0,
+      "tokens_per_correct_answer": 841.2761904761905,
+      "context_compression_ratio": 383.61438435374146,
+      "input_tokens_p50": 707.9571428571429,
+      "input_tokens_p95": 707.9571428571429,
+      "write_cost": 712.7428571428571
+    },
+    "multi_session": {
+      "recall_at_5": 0.33530110307941985,
+      "context_precision": 0.0,
+      "context_recall": 0.0,
+      "answer_relevance": 0.0,
+      "tokens_per_correct_answer": 1321.5147869674186,
+      "context_compression_ratio": 563.5936404757483,
+      "input_tokens_p50": 751.5488721804511,
+      "input_tokens_p95": 751.5488721804511,
+      "write_cost": 757.8195488721805
+    },
+    "temporal_reasoning": {
+      "recall_at_5": 0.2174042228892513,
+      "context_precision": 0.0,
+      "context_recall": 0.0,
+      "answer_relevance": 0.0,
+      "tokens_per_correct_answer": 1931.9671679197995,
+      "context_compression_ratio": 220.12139309953702,
+      "input_tokens_p50": 737.4135338345865,
+      "input_tokens_p95": 737.4135338345865,
+      "write_cost": 748.1203007518797
+    },
+    "knowledge_update": {
+      "recall_at_5": 0.2443019943019943,
+      "context_precision": 0.0,
+      "context_recall": 0.0,
+      "answer_relevance": 0.0,
+      "tokens_per_correct_answer": 1074.474358974359,
+      "context_compression_ratio": 465.1351247517914,
+      "input_tokens_p50": 725.1538461538462,
+      "input_tokens_p95": 725.1538461538462,
+      "write_cost": 730.0128205128206
+    },
+    "single_session_assistant": {
+      "recall_at_5": 0.04176587301587302,
+      "context_precision": 0.0,
+      "context_recall": 0.0,
+      "answer_relevance": 0.0,
+      "tokens_per_correct_answer": 1261.517857142857,
+      "context_compression_ratio": 239.64193247896506,
+      "input_tokens_p50": 775.2857142857143,
+      "input_tokens_p95": 775.2857142857143,
+      "write_cost": 784.3392857142857
+    }
+  }
 }