@warmdrift/kgauto-compiler 2.0.0-alpha.7 → 2.0.0-alpha.70

package/dist/profiles.js

@@ -21,12 +21,27 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 var profiles_exports = {};
 __export(profiles_exports, {
   ALIASES: () => ALIASES,
+  LATENCY_TIER_MS: () => LATENCY_TIER_MS,
+  _setProfileBrainHook: () => _setProfileBrainHook,
   allProfiles: () => allProfiles,
+  allProfilesRaw: () => allProfilesRaw,
   getProfile: () => getProfile,
+  latencyTierOf: () => latencyTierOf,
   profilesByProvider: () => profilesByProvider,
   tryGetProfile: () => tryGetProfile
 });
 module.exports = __toCommonJS(profiles_exports);
+var LATENCY_TIER_MS = {
+  fast: 4e3,
+  medium: 11e3,
+  slow: 24e3
+};
+function latencyTierOf(profile) {
+  if (profile.latencyTier) return profile.latencyTier;
+  if (profile.weaknesses.includes("latency")) return "slow";
+  if (profile.strengths.includes("speed")) return "fast";
+  return "medium";
+}
 var ANTHROPIC_LOWERING_BASE = {
   system: { mode: "inline" },
   cache: {
@@ -49,6 +64,115 @@ var GOOGLE_LOWERING_BASE = {
 };
 var PROFILES_RAW = [
   // ── Anthropic ──
+  {
+    // VERIFIED 2026-06-21 against the claude-api reference. Claude Fable 5 is
+    // Anthropic's most capable model — a NEW tier ABOVE the opus/sonnet/haiku
+    // line (the id has no MAJOR-MINOR version, just `-5`) at 2× Opus pricing
+    // ($10/$50 vs $5/$25). Hand-authored, not cloned: the onboarder regex can't
+    // match this shape (new family + single-segment version), AND a clone would
+    // carry the wrong (opus) pricing. status:'preview' — no brain evidence yet;
+    // it earns per-archetype placement through the machinery like any model, and
+    // at 2× Opus the quality-floor + cost gates do real work before it leads.
+    // API quirk (claude-api ref): an explicit `thinking:{type:"disabled"}` 400s
+    // on Fable 5 (omit instead) — MOOT here: ANTHROPIC_LOWERING_BASE carries no
+    // `thinking` field, so kgauto never sends the param. (L-049/L-081.)
+    id: "claude-fable-5",
+    verifiedAgainstDocs: "2026-06-21",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 10,
+    costOutputPer1m: 50,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
+    weaknesses: ["cost", "latency"],
+    notes: "Most capable Anthropic model (2026-06); a new tier ABOVE Opus at 2\xD7 pricing ($10/$50). 1M ctx, 128k out. Verified 2026-06-21 against the claude-api reference. status:preview \u2014 no brain evidence yet; earns placement via the machinery (cost + quality-floor gates govern). latencyTier derives `slow` (weaknesses includes latency \u2014 minutes-long turns on hard tasks). The thinking-disabled-400 API quirk is moot for kgauto (Anthropic lowering emits no thinking param).",
+    // archetypePerf cloned from claude-opus-4-8 as a conservative estimate
+    // (Fable >= Opus in capability, but unmeasured — no fabricated superiority).
+    // Re-rank once brain rows exist.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      summarize: 8,
+      classify: 8
+    }
+  },
+  {
+    // Auto-onboarded 2026-06-04 from `claude-opus-4-7`; VERIFIED 2026-06-21
+    // against the claude-api reference. Opus 4.8 shares 4.7's surface exactly —
+    // 1M ctx, 128k out, $5/$25, no new breaking changes — so the clone's
+    // capability data was accurate (confirmed field-by-field). archetypePerf is
+    // still cloned from 4.7 (a sound estimate: 4.8 >= 4.7) — re-rank once brain
+    // rows exist. Promoted to status:'current' 2026-06-21 (Sacha: "include Claude
+    // Opus 4.8") — now live-selectable. Ties with the still-current claude-opus-4-7
+    // (identical $5/$25 + cloned perf) resolve to 4-8 by array order. (L-049/L-081.)
+    id: "claude-opus-4-8",
+    verifiedAgainstDocs: "2026-06-21",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 5,
+    costOutputPer1m: 25,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
+    weaknesses: ["cost", "latency"],
+    notes: "Frontier Opus (2026-06), the current recommended Opus-tier default. Verified 2026-06-21 against the claude-api reference: 1M ctx, 128k out, $5/$25 \u2014 identical surface to 4.7, no new breaking changes. Promoted to status:current 2026-06-21 (live-selectable); coexists with claude-opus-4-7 (still current).",
+    // Cloned archetypePerf from claude-opus-4-7 — re-evaluate once brain rows exist.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      summarize: 8,
+      classify: 8
+    }
+  },
   {
     id: "claude-opus-4-7",
     verifiedAgainstDocs: "2026-05-08",
@@ -79,7 +203,24 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
     weaknesses: ["cost", "latency"],
-    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output."
+    notes: "Frontier (2026-05). Step-change improvement over 4.6 in agentic coding. Adaptive thinking only \u2014 no extended-thinking toggle. 1M context, 128k max output.",
+    // Frontier perf. Drops on archetypes where parallel-tool throughput
+    // (hunt) or low-budget cost-sensitivity (classify/summarize) matters
+    // more than reasoning depth.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // strong but Flash dominates parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype; cost-out of frontier
+      classify: 8
+      // overkill; brain-validated cheaper models cover this
+    }
   },
   {
     id: "claude-opus-4-6",
@@ -111,7 +252,20 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "long_context", "reliable_tool_use", "structured_output", "extended_thinking"],
     weaknesses: ["cost", "latency"],
-    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only)."
+    notes: "Predecessor to 4.7. Still current in Anthropic legacy table. Same pricing as 4.7 \u2014 choose 4.7 unless you need extended-thinking budget control (4.7 is adaptive-only).",
+    // One notch below 4.7 across the board — extended-thinking edge does
+    // not flip any archetype ranking. Legacy: chains should prefer 4.7.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      summarize: 8,
+      classify: 8
+    }
   },
   {
     id: "claude-sonnet-4-6",
@@ -135,7 +289,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
     weaknesses: [],
-    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output."
+    notes: "Workhorse. Best price/quality for most multi-turn agentic work. 1M context, 64k max output.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   },
   {
     id: "claude-haiku-4-5",
@@ -165,7 +335,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "cost", "classification", "cache_friendly", "extended_thinking"],
     weaknesses: ["complex_reasoning", "large_tool_sets"],
-    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`."
+    notes: "Cheapest Anthropic. Great for classify, summarize, ask shapes. 200k context, 64k max output. API alias `claude-haiku-4-5` resolves to dated snapshot `claude-haiku-4-5-20251001`.",
+    // Tier 1 cross-provider anchor for short-output chains (classify/
+    // summarize/extract/transform). Falls off on plan/critique where
+    // reasoning depth matters; competes with Pro on cost+latency.
+    archetypePerf: {
+      classify: 8,
+      summarize: 8,
+      ask: 7,
+      transform: 7,
+      extract: 7,
+      hunt: 6,
+      // tool reliability drops at 16 — cliff guard fires
+      generate: 6,
+      plan: 5,
+      critique: 4
+      // reasoning depth gap vs Sonnet/Opus
+    }
   },
   // ── Google ──
   {
@@ -243,7 +429,131 @@ var PROFILES_RAW = [
     ],
     strengths: ["speed", "volume", "classification", "1m_context", "cost"],
     weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
-    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs."
+    notes: "Fast and cheap with 1M context. Quality cliffs at 8K context and 20 tools \u2014 guard with cliffs.",
+    // Master plan §6.2 anchor. Tier 0 for hunt (parallel tool throughput
+    // 15-75 calls/step beats Sonnet — L-040), summarize, classify.
+    archetypePerf: {
+      hunt: 9,
+      // L-040: parallel tool throughput 15-75/step
+      classify: 7,
+      // brain-validated, 218 rows
+      summarize: 7,
+      // brain-validated; cliff strips tools when present
+      transform: 7,
+      ask: 7,
+      generate: 6,
+      plan: 5,
+      extract: 6,
+      // alpha.8 MAX_TOKENS history on structured output
+      critique: 4
+      // reasoning shallower than Sonnet/Opus
+    }
+  },
+  {
+    // ── Gemini 2.5 Flash-Lite ──
+    // Onboarded 2026-05-13 (s22) after the model-release watcher surfaced
+    // it as a UNREGISTERED + NEW candidate. Released by Google July 2025,
+    // stable. Positioned BELOW Flash on the cost/perf frontier:
+    //   input  $0.10/M (Flash $0.30/M)   — 3× cheaper
+    //   output $0.40/M (Flash $2.50/M)   — 6× cheaper
+    //   cache  $0.01/M                    — 1/10 of input (vs Flash 0.25 discount)
+    // Cliffs are HYPOTHESIZED from Flash's known failure modes — Flash-Lite
+    // is a smaller sibling, so we inherit Flash's cliffs at equal-or-tighter
+    // thresholds. The brain will validate/relax these as evidence accumulates
+    // per (archetype, model) tuple. Currently ZERO brain rows for this model.
+    id: "gemini-2.5-flash-lite",
+    verifiedAgainstDocs: "2026-05-13",
+    provider: "google",
+    status: "current",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash: quality degrades above ~8K. Smaller model \u2014 likely degrades faster. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Flash-Lite is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking enabled per Google API (thinking: true). Same drain risk as Flash \u2014 thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact, kgauto commit 3872832). Flash-Lite shares the
+        // same architectural family — almost certainly inherits this cliff.
+        // Ship the guard preemptively; brain telemetry confirms or relaxes.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.1,
+    costOutputPer1m: 0.4,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google's spec is $0.01/M cache vs
+      // $0.10/M input. Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["lowest_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "Bottom-frontier anchor on cost: $0.10/$0.40 per 1M tokens, 1M context, 65K max output. Released July 2025 (stable). Positioned for classify / summarize / transform archetypes where quality bar is forgiving. Cliffs inherited from Flash at equal-or-tighter thresholds \u2014 re-tune per (archetype) once brain has n\u226520 rows. Alpha.8 contract layer handles MAX_TOKENS-on-structured-output via fallback chain, so structuredOutput=native is safe to declare even though Flash had alpha.8 history. Cache discount in spec: $0.01/M = 1/10 of input (richer than Flash 25%) \u2014 meaningful for repeat-prompt workloads.",
+    // Tier 3 emergency floor for summarize/classify chains. ZERO brain
+    // rows — all values are starter hypotheses anchored to "smaller
+    // sibling of Flash, at-or-below Flash perf on every archetype." The
+    // first 50 brain rows per archetype will validate or relax these.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
   },
   {
     id: "gemini-2.5-pro",
@@ -279,7 +589,21 @@ var PROFILES_RAW = [
       }
     ],
     strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
-    weaknesses: ["pricing_above_200k"]
+    weaknesses: ["pricing_above_200k"],
+    // Master plan §3.3 anchor: tier-2 cross-provider in almost every chain.
+    // Sits on the frontier at perf-9 — close to Sonnet but cheaper input.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      ask: 8,
+      generate: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 8,
+      // tier 1 cross-provider for hunt chain
+      summarize: 7,
+      classify: 7
+    }
   },
   {
     id: "gemini-3.1-pro-preview",
@@ -317,7 +641,23 @@ var PROFILES_RAW = [
     ],
     strengths: ["reasoning", "1m_context", "agentic_coding", "structured_output", "tool_use"],
     weaknesses: ["cost", "preview_status", "pricing_above_200k"],
-    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA."
+    notes: "Frontier Gemini (preview, 2026-Q2). Step-change agentic coding per Google. Cache discount 10\xD7 (vs 4\xD7 for 2.5 Pro). Use status=preview to flag rollback path until GA.",
+    // Frontier-Gemini preview — bumped one notch over 2.5 Pro on agentic
+    // coding / reasoning per Google's release notes. Preview status:
+    // chains should stay on 2.5 Pro until GA. Starter hypothesis.
+    archetypePerf: {
+      critique: 10,
+      // Google claims step-change on reasoning
+      plan: 10,
+      ask: 9,
+      generate: 9,
+      extract: 9,
+      transform: 8,
+      hunt: 9,
+      // step-change agentic per Google
+      summarize: 8,
+      classify: 7
+    }
   },
   // ── DeepSeek ──
   // 2026-05-08 audit (L-073): DeepSeek's `deepseek-chat` was silently aliased
@@ -357,11 +697,50 @@ var PROFILES_RAW = [
     ],
     strengths: ["cost", "1m_context", "json_output", "code", "reasoning"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES."
+    // alpha.47 — explicit slow override. Tag derivation would say 'medium'
+    // (no 'latency' weakness, no 'speed' strength), but the alpha.46 shadow
+    // probe MEASURED deepseek-v4-flash at 20485ms served (2026-06-03) — ~2.3×
+    // gemini-2.5-flash on PB's synchronous /api/analyze path. The 'flash' name
+    // is DeepSeek's, not a speed promise. This is the row that, scoring 0.85
+    // baseQuality (it carries 'reasoning') with no latency counterweight,
+    // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
+    latencyTier: "slow",
+    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES.",
+    // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
+    // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
+    // Falls off on hunt (sequential tools — L-040) and reasoning depth.
+    archetypePerf: {
+      classify: 7,
+      // brain-validated, 169 rows
+      summarize: 7,
+      // archetype-tolerant, no brain evidence yet
+      ask: 6,
+      transform: 6,
+      generate: 5,
+      plan: 5,
+      extract: 5,
+      critique: 4,
+      hunt: 4
+      // sequential tool calls only — L-040
+    },
+    // alpha.43 — family-level conventions for the `deepseek-chat` family
+    // (V4-Flash + future non-thinking-mode members). V4-Flash struggles
+    // with summarize-archetype + structured-output: 3 of 4 schema-attempted
+    // PB calls (2026-05-18 + 2026-05-22) hit `structured_output_parse_failed`.
+    // Surface the friction at compile time so consumers either drop the
+    // schema or route to a more schema-reliable family.
+    archetypeConventions: [
+      {
+        archetype: "summarize",
+        structuredOutputHint: "avoid",
+        cliffWarning: "deepseek-chat family schema compliance on summarize is weak (3 of 4 schema-attempted PB calls hit structured_output_parse_failed). Consider routing summarize+structured-output to a different family (claude-sonnet, gemini-pro, openai-gpt-5.4) or dropping the schema for free-text summaries.",
+        reason: "Brain evidence: PB compile_outcomes 2026-05-18 + 2026-05-22, error_type=structured_output_parse_failed on deepseek-v4-flash + summarize archetype."
+      }
+    ]
   },
   {
     id: "deepseek-v4-pro",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-06-22",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -379,10 +758,14 @@ var PROFILES_RAW = [
         reason: "Sequential tool calls only \u2014 L-040"
       }
     ],
-    // Profile carries REGULAR pricing, not the 75%-off promo (ends 2026-05-31).
-    // Under-estimating cost is worse than over-estimating for budget caps.
-    costInputPer1m: 1.74,
-    costOutputPer1m: 3.48,
+    // Verified against the live DeepSeek pricing page 2026-06-22: the 75%-off
+    // launch "promo" ($0.435/$0.87) did NOT revert on 2026-05-31 — it became the
+    // standing price (docs show $0.435 in / $0.87 out cache-miss, $0.003625
+    // cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
+    // was a projected revert that never happened (L-073/L-081) — it 4×-overstated
+    // cost and suppressed V4-Pro in cost-aware ranking.
+    costInputPer1m: 0.435,
+    costOutputPer1m: 0.87,
     lowering: {
       system: { mode: "inline" },
       cache: { strategy: "unsupported" },
@@ -393,7 +776,825 @@ var PROFILES_RAW = [
     ],
     strengths: ["quality", "reasoning", "1m_context", "json_output", "code", "extended_thinking"],
     weaknesses: ["parallel_tools", "large_tool_sets"],
-    notes: "Pro tier. 1M context, 384k max output. Regular pricing $1.74/$3.48; 75% promo through 2026-05-31 ($0.435/$0.87). Default mode = thinking."
+    // alpha.47 — explicit slow override. Measured 47722ms on the alpha.46
+    // shadow probe (2026-06-03) — it's an extended-thinking reasoner, slowest
+    // of the served set. Tag derivation would say 'medium'; the measurement says
+    // otherwise.
+    latencyTier: "slow",
+    notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking.",
+    // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
+    // bumped one notch over V4-Flash; same parallel-tool ceiling.
+    archetypePerf: {
+      plan: 7,
+      // §3.3 tier 3 for plan
+      critique: 6,
+      ask: 7,
+      generate: 6,
+      classify: 7,
+      summarize: 7,
+      extract: 6,
+      transform: 6,
+      hunt: 4
+      // sequential tools — same as V4-Flash
+    },
+    // alpha.43 — family-level conventions for the `deepseek-reasoner` family
+    // (V4-Pro is the family rep; future thinking-mode members inherit).
+    //
+    // The decisive-archetype suffix is the load-bearing piece: 2026-05-28
+    // first real V4-Pro probe on tt-intel/classify (exclusion-finding ID 20)
+    // showed 8/10 judge rationales citing "candidate hedges and fails to
+    // commit to a single classification" or "candidate refuses to pick
+    // among the labels." Without the forcing-function suffix, every
+    // reasoner probe on classify/extract/ask verdicts stay-excluded for
+    // model-behavior reasons (epistemic humility is reasoner-shaped),
+    // NOT quality reasons. The suffix re-shapes the contract so the model
+    // is asked for what it's actually being measured on.
+    //
+    // The hunt-archetype cliffWarning is the L-040 architectural reading
+    // surfaced at compile time. Reasoners produce sequential reasoning
+    // chains by design; hunt archetypes with 3+ parallel tools lose
+    // substantial throughput (L-040 + L-041). The L-040 cliff in the
+    // `cliffs` array already trims to top-relevant tools; this warning
+    // tells the consumer the family is wrong for the shape so they can
+    // route to gemini-flash (parallel-strong) instead of accepting trimmed
+    // throughput silently.
+    archetypeConventions: [
+      {
+        archetype: "classify",
+        promptSuffix: "\n\nOutput exactly one classification label. Do not hedge, do not add preamble, do not explain your reasoning. If uncertain, pick the most likely label and commit to it.",
+        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: 2026-05-28 V4-Pro probe on tt-intel/classify (exclusion-finding ID 20, brain probe_outcomes); 8 of 10 judge rationales cited hedging or refusal-to-commit. Forcing-function suffix converts reasoner output to decisive labels."
+      },
+      {
+        archetype: "extract",
+        promptSuffix: "\n\nOutput exactly the requested structured fields. Do not hedge, do not add preamble, do not explain your reasoning. If a field is unknown, emit the field with a null/empty value and commit to that \u2014 do not refuse to extract.",
+        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: same family failure shape as classify (exclusion-finding ID 20, 2026-05-28). Forcing-function suffix asks for commitment instead of epistemic-humility hedging."
+      },
+      {
+        archetype: "ask",
+        promptSuffix: "\n\nAnswer the question directly. Do not hedge, do not add preamble, do not explain your reasoning unless explicitly asked. If uncertain, give your best answer and commit to it.",
+        reason: "Reasoners (deepseek-reasoner family) hedge on decisive-output archetypes by design. Evidence: same family failure shape as classify (exclusion-finding ID 20, 2026-05-28). Forcing-function suffix asks for commitment instead of epistemic-humility hedging."
+      },
+      {
+        archetype: "hunt",
+        whenToolCountAtLeast: 3,
+        cliffWarning: "Reasoners (deepseek-reasoner family) produce sequential reasoning chains by design; hunt archetypes with 3+ parallel tools lose substantial throughput. Consider gemini-flash family (parallel-strong) instead \u2014 see L-040 / L-041.",
+        reason: "L-040 + L-041 architectural reading; 2026-05-28 V4-Pro probe evidence (exclusion-finding ID 20) confirms reasoner-family hunt fit is structurally weak."
+      }
+    ]
+  },
+  // ── OpenAI ──
+  // alpha.16 (2026-05-17): close the half-supported provider gap. env.ts
+  // already registered OPENAI_API_KEY + executeOpenAI + normalizeOpenAILike
+  // + lowerOpenAI all existed; profile entries were missing, so the
+  // alpha.10 auto-filter would mark openai-keyed models reachable but
+  // there were no profiles to filter IN. Half-supported is now fully
+  // supported. PB request `openai-provider-profiles` (2026-05-16).
+  //
+  // Profile data verified against developers.openai.com/api/docs/pricing
+  // + per-model pages 2026-05-17. L-049/L-081 step-zero: no AI-trained
+  // numbers — fetched live from OpenAI's docs. As of 2026-05, OpenAI's
+  // current flagship is gpt-5.5 (2025-12 cutoff); gpt-5.4-{base,mini,nano}
+  // are the workhorse family. gpt-4.1 + gpt-4o are legacy.
+  //
+  // Both 5.5 and 5.4 carry a 272K input-token pricing cliff (2x input,
+  // 1.5x output beyond that). Modeled as a `downgrade_quality_warning`
+  // cliff because it ranks the model down at large-context shapes — the
+  // semantics of "this model is now 2x more expensive" map onto the
+  // existing penalty mechanism. Cost-watcher will catch high-context
+  // spikes empirically; the cliff prevents naive routing into the doubled
+  // pricing zone.
+  {
+    id: "gpt-5.5",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 5,
+    costOutputPer1m: 30,
+    lowering: {
+      system: { mode: "inline" },
+      // OpenAI caching is implicit (auto-applied to repeated prefixes
+      // ≥1024 tokens for prompt_tokens_details.cached_tokens). No
+      // wire-format marker. Discount: 10x for cached input ($0.50/$5.00).
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "structured_output", "reliable_tool_use", "reasoning_effort_knob"],
+    weaknesses: ["cost", "pricing_cliff_at_272k"],
+    notes: "OpenAI frontier (2026-05). 1M context (1.05M total), 128K max output, 2025-12 cutoff. Reasoning effort knob (none/low/medium/high/xhigh). Pricing cliff at 272K input.",
+    // Frontier-tier perf hypothesis. Anchored to Opus 4.7 row (similar
+    // price/positioning). Brain evidence will refine; no telemetry yet.
+    archetypePerf: {
+      critique: 9,
+      plan: 9,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      // parallel tool support good but cliff at 272K hurts deep multi-step
+      summarize: 7,
+      // overkill for tolerant archetype
+      classify: 7
+      // overkill; cheaper models cover this
+    }
+  },
+  {
+    id: "gpt-5.4",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 105e4,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 272e3,
+        action: "downgrade_quality_warning",
+        reason: "OpenAI pricing tier shift: >272K input tokens billed at 2x input + 1.5x output rates"
+      }
+    ],
+    costInputPer1m: 2.5,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["reasoning", "long_context", "structured_output", "reliable_tool_use"],
+    weaknesses: ["pricing_cliff_at_272k"],
+    notes: "OpenAI workhorse (2026-05). 1M context (1.05M total), 128K max output, 2025-08 cutoff. Pricing cliff at 272K input. Pairs cleanly with Sonnet 4.6 on cost ($2.50/$15.00 vs $3.00/$15.00).",
+    // Anchored to Sonnet 4.6 row (similar price/positioning). Slight
+    // anthropic-side edge on agentic coding per master plan vibe.
+    archetypePerf: {
+      critique: 8,
+      plan: 8,
+      generate: 8,
+      ask: 8,
+      extract: 8,
+      transform: 8,
+      hunt: 7,
+      summarize: 7,
+      classify: 7
+    }
+  },
+  {
+    id: "gpt-5.4-mini",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.75,
+    costOutputPer1m: 4.5,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "agentic_coding", "structured_output", "reliable_tool_use"],
+    weaknesses: ["reasoning_depth"],
+    notes: "OpenAI mini-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. OpenAI describes as 'strongest mini model for coding, computer use, subagents.' Cache discount 10x ($0.075 input).",
+    // Mini-tier hypothesis. Anchored to Haiku 4.5 + Flash row pricing.
+    // Cost is slightly higher than Haiku ($0.75 vs $0.50 input) but
+    // OpenAI claims strong coding/subagent perf.
+    archetypePerf: {
+      ask: 7,
+      generate: 7,
+      extract: 7,
+      transform: 7,
+      classify: 7,
+      summarize: 7,
+      hunt: 7,
+      plan: 6,
+      critique: 5
+      // reasoning depth gap — frontier models handle this
+    }
+  },
+  {
+    id: "gpt-5.4-nano",
+    verifiedAgainstDocs: "2026-05-17",
+    provider: "openai",
+    status: "current",
+    maxContextTokens: 4e5,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 0.2,
+    costOutputPer1m: 1.25,
+    lowering: {
+      system: { mode: "inline" },
+      cache: { strategy: "unsupported", minTokens: 1024, discount: 0.1 },
+      tools: { format: "openai" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate to fallback chain" },
+      { signal: "model_not_found", action: "escalate", reason: "Model deprecated/renamed \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "speed", "volume", "structured_output"],
+    weaknesses: ["reasoning_depth", "no_computer_use"],
+    notes: "OpenAI nano-tier (2026-05). 400K context, 128K max output, 2025-08 cutoff. 'Cheapest GPT-5.4-class for simple high-volume tasks.' No fine-tuning, no computer-use tools. Cache discount 10x.",
+    // Nano-tier. Anchored to Flash-Lite row ($0.10/$0.40 vs nano's
+    // $0.20/$1.25). Slightly more expensive than Flash-Lite but with
+    // OpenAI brand reliability. Good fit for classify/summarize floor.
+    archetypePerf: {
+      classify: 7,
+      summarize: 6,
+      ask: 6,
+      transform: 6,
+      extract: 6,
+      generate: 5,
+      hunt: 5,
+      plan: 4,
+      critique: 3
+      // not for reasoning archetypes
+    }
+  },
+  // ── Auto-onboarded (UNVERIFIED) ──
+  // Cloned by scripts/auto-onboard-models.mjs from a same-family template.
+  // Each entry's pricing/context/cliffs/lowering reflects the template, NOT
+  // provider docs. Verify before promoting status to 'current' (L-049/L-081).
+  {
+    // s37 (2026-05-21): UNVERIFIED-AUTO-ONBOARD → verified against
+    // ai.google.dev/gemini-api/docs/models/gemini-3-flash-preview +
+    // ai.google.dev/gemini-api/docs/pricing. L-081 catches:
+    //   maxOutputTokens 65_535 → 65_536 (off-by-one)
+    //   costInputPer1m  0.30 → 0.50  (template-cloned from 2.5-flash; actual is 1.67× more expensive)
+    //   costOutputPer1m 2.50 → 3.00  (template-cloned; actual 1.2× more expensive)
+    //   cache discount  default 0.25 → 0.10 (10× discount, $0.05/$0.50 per docs)
+    // Cliffs inherited from 2.5-flash conservatively. The 8K-context-quality
+    // cliff was a 2.5-Flash observation — Google positions Gemini 3 as
+    // sustained-frontier-on-long-context; brain evidence will validate/relax.
+    // Kept as guard for now.
+    id: "gemini-3-flash-preview",
+    verifiedAgainstDocs: "2026-05-21",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from 2.5-flash guard; brain evidence on Gemini 3 long-context quality will validate/relax"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Tool reliability drops above ~20 tools (despite 128 hard limit) \u2014 inherited from Flash family"
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking tokens consume maxOutputTokens \u2014 empty response if drained"
+      },
+      {
+        // Inherited from gemini-2.5-flash s11 trust artifact. Family-likely
+        // failure mode for Flash architecture. Keep preemptively until brain
+        // evidence on Gemini 3 specifically.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on 3-flash-preview specifically."
+      }
+    ],
+    costInputPer1m: 0.5,
+    costOutputPer1m: 3,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // 10× cache discount per Google pricing: $0.05/M cached vs $0.50/M input.
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known: empty after tool result \u2014 retry with thinking off"
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off"
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target"
+      }
+    ],
+    strengths: ["speed", "volume", "classification", "1m_context", "cost"],
+    weaknesses: ["complex_schemas", "large_tool_sets", "high_context_quality"],
+    notes: "Verified s37 (2026-05-21) against Google docs. Step-change positioning vs 2.5-flash on agentic loops per Google's release notes (Dec 2025). Pricing 1.67\xD7/1.2\xD7 higher than 2.5-flash; cache discount 10\xD7 (vs 4\xD7 for 2.5). Status=preview until brain evidence accumulates.",
+    // Anchored to 2.5-flash archetypePerf as starter, with judgment adjustments
+    // for Google's "step-change on agentic" positioning. Brain evidence (zero
+    // rows today) will replace these starter values.
+    archetypePerf: {
+      hunt: 9,
+      // Inherits 2.5-flash L-040 parallel-tool tier; Google positions 3 as agentic-loop upgrade
+      classify: 7,
+      // Inherits 2.5-flash brain-validated tier (218 rows on 2.5)
+      summarize: 7,
+      // Inherits 2.5-flash; cliff strips tools when present
+      transform: 7,
+      ask: 8,
+      // +1 vs 2.5-flash — sustained-frontier positioning
+      generate: 7,
+      // +1 vs 2.5-flash — agentic coding upgrade per Google
+      plan: 6,
+      // +1 vs 2.5-flash — complex iterations per positioning
+      extract: 6,
+      critique: 5
+      // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
+    }
+  },
+  {
+    // ── Gemini 3.1 Flash-Lite ──
+    // Onboarded 2026-05-16 by auto-onboarder; s37 (2026-05-21) verified
+    // against ai.google.dev/gemini-api/docs/pricing.
+    //
+    // L-081 CATCHES (template clone from 2.5-flash-lite was 2.5-3.75× too cheap):
+    //   costInputPer1m  0.10 → 0.25 (template clone undervalued by 2.5×)
+    //   costOutputPer1m 0.40 → 1.50 (template clone undervalued by 3.75×)
+    //
+    // Real 3.1-flash-lite is NOT a cost-equivalent successor to 2.5-flash-lite —
+    // it sits between 2.5-flash-lite ($0.10/$0.40) and 2.5-flash ($0.30/$2.50).
+    // Cache discount 10× verified ($0.025/M cached vs $0.25/M input).
+    //
+    // Cliffs are HYPOTHESIZED from 2.5-flash family; brain evidence pending.
+    id: "gemini-3.1-flash-lite",
+    verifiedAgainstDocs: "2026-05-21",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from Flash: quality degrades above ~8K. Smaller model \u2014 likely degrades faster. Re-tune from brain after n\u226520."
+      },
+      {
+        metric: "tool_count",
+        threshold: 10,
+        action: "drop_to_top_relevant",
+        reason: "Conservative: Flash drops at 20, Flash-Lite is smaller \u2014 assume tighter ceiling until brain proves otherwise."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking enabled per Google API (thinking: true). Same drain risk as Flash \u2014 thinking tokens consume maxOutputTokens."
+      },
+      {
+        // Strong prior: Flash hit 5/5 empty rate on summarize+tools (s11
+        // trust artifact, kgauto commit 3872832). Flash-Lite shares the
+        // same architectural family — almost certainly inherits this cliff.
+        // Ship the guard preemptively; brain telemetry confirms or relaxes.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from Flash s11 cliff: summarize+tools \u2192 empty response. Preemptive guard until brain evidence on Flash-Lite specifically."
+      }
+    ],
+    costInputPer1m: 0.25,
+    costOutputPer1m: 1.5,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // Cache discount 10× (vs Flash 4×) — Google docs s37: $0.025/M cached vs
+      // $0.25/M input. Material for repeat-prompt workloads (classify shape).
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Known on Flash family: empty after tool result \u2014 retry with thinking off."
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off."
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target."
+      }
+    ],
+    strengths: ["low_cost", "speed", "volume", "classification", "summarize", "1m_context", "cache_friendly"],
+    weaknesses: ["complex_reasoning", "large_tool_sets", "complex_schemas", "structured_output_unproven", "long_context_quality"],
+    notes: "Verified s37 (2026-05-21) against Google docs. Sits between 2.5-flash-lite (cheaper) and 2.5-flash (more expensive) on cost frontier; 2.5\xD7 more expensive than initial template-clone. Cliffs hypothesized from Flash family \u2014 brain evidence pending.",
+    // Tier 2-3 floor for summarize/classify chains at the new (verified) price
+    // point. ZERO brain rows — values are starter hypotheses anchored to
+    // "smaller sibling of Flash at higher cost than 2.5-flash-lite." The first
+    // 50 brain rows per archetype will validate or relax these.
+    archetypePerf: {
+      classify: 6,
+      // starter hypothesis — verify (Flash is 7, lite likely ≤)
+      summarize: 6,
+      // starter hypothesis — verify; cliff strips tools
+      transform: 6,
+      // starter hypothesis — verify
+      ask: 5,
+      hunt: 5,
+      generate: 4,
+      extract: 4,
+      plan: 3,
+      critique: 3
+    }
+  },
+  // ─────────────────────────────────────────────────────────────────────────
+  // Gemini 3.5 Flash — hand-onboarded s37 (2026-05-21)
+  //
+  // Google positioning ("Most intelligent for sustained frontier performance
+  // on agentic and coding tasks" / "particularly effective for rapid agentic
+  // loops involving complex coding cycles and iterations") suggests this is
+  // the Flash-family upgrade specifically aimed at hunt-shape workloads.
+  // Pricing 5× input / 3.6× output vs 2.5-flash — material cost premium.
+  // archetypePerf adjusted +1 vs 2.5-flash on ask/generate/plan/critique
+  // (sustained-frontier positioning); hunt held at 9 inherited from L-040
+  // family parallel-tool tier; brain evidence will validate within 50 rows.
+  //
+  // Cliffs inherited conservatively from 2.5-flash. Google's "sustained
+  // frontier on long-context" positioning suggests the 8K cliff may not
+  // apply to 3.5 — keep as guard until brain evidence shows otherwise.
+  //
+  // Specs verified against:
+  //   ai.google.dev/gemini-api/docs/models/gemini-3.5-flash
+  //   ai.google.dev/gemini-api/docs/pricing (Standard tier)
+  // ─────────────────────────────────────────────────────────────────────────
+  {
+    id: "gemini-3.5-flash",
+    verifiedAgainstDocs: "2026-05-21",
+    provider: "google",
+    status: "current",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited from 2.5-flash guard; Google positions 3.5 as sustained-frontier-on-long-context but brain evidence pending"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Inherited from Flash family: tool reliability drops above ~20 (despite 128 hard limit). Validate per (archetype, model) after n\u226520."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking mode supported per Google docs; same drain risk as 2.5-flash \u2014 thinking tokens consume maxOutputTokens"
+      },
+      {
+        // Inherited from 2.5-flash s11 trust artifact (5/5 empty rate on
+        // tt-intelligence/summarize/gemini-2.5-flash with tools offered).
+        // Family-likely failure mode for Flash architecture across versions.
+        // Keep preemptively until brain evidence on 3.5-flash specifically.
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff (kgauto commit 3872832): summarize+tools \u2192 empty response. Preemptive guard until brain evidence on 3.5-flash specifically."
+      }
+    ],
+    costInputPer1m: 1.5,
+    costOutputPer1m: 9,
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      // 10× cache discount per Google pricing: $0.15/M cached vs $1.50/M input.
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Inherited Flash-family pattern: empty after tool result \u2014 retry with thinking off"
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off"
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target"
+      }
+    ],
+    strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode", "sustained_frontier"],
+    weaknesses: ["cost_vs_2_5_flash", "no_brain_evidence_yet"],
+    notes: "Hand-onboarded s37 (2026-05-21) verified against Google docs. Stable status; positioned as Flash-family upgrade for agentic loops and coding. 5\xD7/3.6\xD7 more expensive than 2.5-flash but Google claims step-change on sustained frontier work. archetypePerf adjustments are judgment-grounded starter hypotheses \u2014 brain evidence will validate within ~50 rows per archetype.",
+    // Starter hypothesis: anchored to 2.5-flash archetypePerf with +1
+    // adjustments where Google's positioning explicitly supports
+    // (agentic/coding/sustained). Hunt held at 9 inherited from L-040 family
+    // parallel-tool tier. Brain evidence will replace.
+    archetypePerf: {
+      hunt: 9,
+      // Inherited from 2.5-flash L-040 parallel-tool tier; Google positions 3.5 as agentic-loop champion
+      classify: 7,
+      // Inherited from 2.5-flash brain-validated tier (218 rows on 2.5)
+      summarize: 7,
+      // Inherited from 2.5-flash; cliff strips tools when present
+      transform: 7,
+      ask: 8,
+      // +1 vs 2.5-flash — sustained-frontier positioning
+      generate: 8,
+      // +1 vs 2.5-flash (6→8) — Google: "complex coding cycles and iterations"
+      plan: 7,
+      // +1 vs 2.5-flash (5→7) — "complex iterations" positioning
+      extract: 7,
+      // +1 vs 2.5-flash — sustained-frontier on structured tasks
+      critique: 5
+      // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
+    }
+  },
+  {
+    // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
+    // against the claude-api reference (cc-portfolio ratification pass). The
+    // clone got context right (1M) and pricing right at sticker ($3/$15 —
+    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
+    // encoded here per the time-bounded-pricing convention, intro belongs in
+    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
+    // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
+    // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
+    // return 400 — moot for kgauto's own call() path (ANTHROPIC_LOWERING_BASE
+    // emits no sampling params) but a REAL hazard for compileForAISDKv6
+    // consumers that pass temperature themselves (tt-intel scoring uses temp:0
+    // for determinism — that 400s on this model; noted to consumers via the
+    // contract). (b) Omitting `thinking` runs ADAPTIVE by default (4-6 ran
+    // thinking-off) — output spend shifts. (c) Supports effort xhigh. status:
+    // 'preview' per the Fable precedent — no brain evidence yet; promotion to
+    // 'current' is an explicit call. (L-049/L-081.)
+    id: "claude-sonnet-5",
+    verifiedAgainstDocs: "2026-07-02",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
+    weaknesses: [],
+    notes: "Sonnet 5 (2026-06): near-Opus quality on coding/agentic work at Sonnet cost. Verified 2026-07-02 against the claude-api reference: 1M ctx, 128k out (clone's 64k corrected), $3/$15 sticker (intro $2/$10 through 2026-08-31). New tokenizer ~30% more tokens vs sonnet-4-6. Consumer hazards: non-default temperature/top_p/top_k 400 (temp:0 rejected \u2014 deterministic-scoring consumers must omit); thinking defaults to adaptive when omitted. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
+  },
+  // ── Z.ai (GLM) ──
+  // alpha.65 (2026-07-18): first Z.ai profile, provider `zai` added same
+  // release. Step-zero sources (L-049/L-081, all fetched 2026-07-18):
+  // docs.z.ai/guides/llm/glm-5.2 (capabilities), docs.z.ai api-reference
+  // chat-completion (wire params: thinking.type enabled|disabled default
+  // enabled, reasoning_effort max..none, tool_stream, max output 131072),
+  // docs.z.ai/guides/overview/pricing ($1.4 in / $4.4 out, cached input
+  // $0.26, cache storage limited-time free), vLLM recipe (native ctx
+  // 1,048,576). NO AI-trained numbers.
+  {
+    id: "glm-5.2",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "zai",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 131072,
+    // No documented tool-count cap — conservative default pending measured
+    // evidence; matches the deepseek posture, not a doc number.
+    maxTools: 16,
+    // Hosted-API docs document tools + tool_choice + tool_stream but say
+    // NOTHING about multiple tool_calls per response. Conservative false so
+    // hunt-shaped ranking doesn't over-promise; flip on doc or probe
+    // evidence (L-081: undocumented ≠ true).
+    parallelToolCalls: false,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 1.4,
+    costOutputPer1m: 4.4,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic prefix caching ("intelligent caching mechanism") — no wire
+      // marker to emit, DeepSeek-style. Cache-hit input $0.26/M (~0.19× of
+      // miss) tracked in notes; storage limited-time free as of 2026-07-18.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" },
+      // thinking.type: 'enabled' (server default) | 'disabled'. lowerZai
+      // emits an explicit disable only on forceThinkingZero cliffs.
+      thinking: { field: "thinking.type", default: "auto" }
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["cost", "1m_context", "code", "tool_use", "json_output", "reasoning"],
+    weaknesses: ["parallel_tools"],
+    notes: "GLM-5.2 (Z.ai, 2026-06): agentic-coding flagship. 1,048,576 ctx / 131,072 max out. $1.40/$4.40 per 1M; cached input $0.26/M (automatic caching, no marker; storage limited-time free as of 2026-07-18). Thinking on by default (`thinking.type`), `reasoning_effort` max..none. Parallel tool calls UNDOCUMENTED on the hosted API \u2014 profile says false until doc or probe evidence. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on the vendor's agentic-coding positioning at a
+    // mid-cost tier, discounted for the unproven parallel-tool story.
+    archetypePerf: {
+      generate: 7,
+      // agentic-coding positioning — starter hypothesis
+      transform: 7,
+      // starter hypothesis
+      plan: 7,
+      // starter hypothesis
+      summarize: 7,
+      // tolerant archetype — starter hypothesis
+      ask: 6,
+      extract: 6,
+      critique: 6,
+      classify: 6,
+      hunt: 5
+      // tools work; parallel unproven → discounted
+    }
+  },
+  // ── Moonshot (Kimi) ──
+  // alpha.65 (2026-07-18): first Moonshot profile, provider `moonshot` added
+  // same release. Step-zero sources (all fetched 2026-07-18):
+  // platform.kimi.ai/docs/pricing/chat-k3 ($3.00 cache-miss in / $0.30
+  // cache-hit in / $15.00 out; 1,048,576 ctx; ToolCalls + JSON Mode +
+  // Partial Mode), kimi-k3-quickstart (max_completion_tokens default
+  // 131072, up to 1,048,576; base https://api.moonshot.ai/v1; vision via
+  // base64 or ms://file-id, NO public image URLs), models-overview
+  // (temperature FIXED 1.0, top_p FIXED 0.95, n FIXED 1, always reasons,
+  // reasoning_effort 'max' only), tool-calls guide ("can choose to call
+  // multiple tools at once … will tend to call them in parallel").
+  {
+    id: "kimi-k3",
+    verifiedAgainstDocs: "2026-07-18",
+    provider: "moonshot",
+    status: "preview",
+    maxContextTokens: 1048576,
+    // Documented ceiling; default is 131,072 when unset.
+    maxOutputTokens: 1048576,
+    // No hard doc cap; docs recommend dynamic tool loading at "dozens or
+    // hundreds" — 64 matches the frontier-agentic posture (sonnet-5 tier).
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: {
+      system: { mode: "inline" },
+      // Automatic context caching — no wire marker. Cache-hit input $0.30/M
+      // (0.1× of miss) tracked in notes.
+      cache: { strategy: "unsupported" },
+      tools: { format: "openai" }
+      // No thinking knob: reasoning is always-on and reasoning_effort
+      // accepts only 'max' (= the server default). Nothing to lower.
+    },
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "reasoning", "1m_context", "parallel_tools", "tool_use", "code", "vision"],
+    // 'latency' → latencyTierOf derives 'slow': always-on max-effort
+    // reasoner with no off switch; no measured p50 yet, the tag is the
+    // honest prior (reasoner family precedent: deepseek-v4-pro ~48s).
+    weaknesses: ["cost", "latency"],
+    notes: "Kimi K3 (Moonshot, API live 2026-07-16): 2.8T-param open-weight flagship, native vision. 1,048,576 ctx; max_completion_tokens default 131,072, ceiling 1,048,576. $3.00/$15.00 per 1M, cache-hit input $0.30/M (automatic caching). Reasoning ALWAYS ON (reasoning_effort max only, cannot disable; responses may carry reasoning_content). Wire quirks: temperature fixed 1.0 / top_p 0.95 / n 1 (kgauto never emits those); vision input must be base64 or ms://file-id \u2014 public image URLs rejected. status:preview \u2014 earns placement via the machinery.",
+    // Starter hypotheses — verify with telemetry/probes; NO brain evidence
+    // yet. Anchored on documented long-horizon-agentic + parallel-tool
+    // strengths; discounted on terse archetypes where always-on reasoning
+    // bills $15/M output for thinking the archetype doesn't need.
+    archetypePerf: {
+      hunt: 8,
+      // parallel tools + long-horizon agentic — starter hypothesis
+      plan: 8,
+      // starter hypothesis
+      generate: 8,
+      // starter hypothesis
+      critique: 7,
+      ask: 7,
+      extract: 7,
+      transform: 7,
+      summarize: 6,
+      // works, but $15/M output reasoning tax on a tolerant archetype
+      classify: 5
+      // always-on reasoning cost + reasoner hedge risk
+    },
+    // Diagnostics-only (no prompt rewrite): reasoner-family hedge risk on
+    // decisive archetypes, carried as a warning until K3-specific probe
+    // evidence exists. The deepseek-reasoner family showed 8/10 judge
+    // rationales citing hedging on classify (exclusion-finding ID 20,
+    // 2026-05-28); K3 is an always-on reasoner and inherits the risk class,
+    // not the conviction.
+    archetypeConventions: [
+      {
+        archetype: "classify",
+        cliffWarning: "kimi-k3 is an always-on reasoner: decisive archetypes (classify) may hedge, and reasoning tokens bill at $15/M output. Family-prior risk (deepseek-reasoner precedent), not K3-measured \u2014 probe before relying on it for classify.",
+        reason: "Reasoner-family prior (exclusion-finding ID 20, 2026-05-28, deepseek-v4-pro on tt-intel/classify). No K3-specific evidence yet \u2014 warning only, no prompt rewrite."
+      }
+    ]
   }
 ];
 var ALIASES = {
@@ -403,14 +1604,20 @@ var ALIASES = {
   // Legacy kgauto typo — actual API alias is dash-form (alpha.1 had dot).
   "claude-haiku-4.5": "claude-haiku-4-5"
 };
+var brainHook = {};
+function _setProfileBrainHook(hook) {
+  brainHook = hook;
+}
 function canonicalId(id) {
-  return ALIASES[id] ?? id;
+  return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
 }
 var PROFILE_INDEX = new Map(
   PROFILES_RAW.map((p) => [p.id, p])
 );
 function getProfile(id) {
   const canonical = canonicalId(id);
+  const fromBrain = brainHook.getProfile?.(canonical);
+  if (fromBrain) return fromBrain;
   const p = PROFILE_INDEX.get(canonical);
   if (!p) {
     const known = [...PROFILE_INDEX.keys(), ...Object.keys(ALIASES)].join(", ");
@@ -419,19 +1626,27 @@ function getProfile(id) {
   return p;
 }
 function tryGetProfile(id) {
-  return PROFILE_INDEX.get(canonicalId(id));
+  const canonical = canonicalId(id);
+  return brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
 }
 function allProfiles() {
   return PROFILES_RAW;
 }
+function allProfilesRaw() {
+  return PROFILES_RAW;
+}
 function profilesByProvider(provider) {
   return PROFILES_RAW.filter((p) => p.provider === provider);
 }
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   ALIASES,
+  LATENCY_TIER_MS,
+  _setProfileBrainHook,
   allProfiles,
+  allProfilesRaw,
   getProfile,
+  latencyTierOf,
   profilesByProvider,
   tryGetProfile
 });