feat(planner): ground EV math in observed usage + persist EV verdict as shadow telemetry

[aminsamir45]

Tier 0 of the inter-turn routing optimization plan: make the planner's
inputs and outputs trustworthy before any switching policy is armed.

1. Observed-token grounding: the EV math priced the prefix with
   feats.Tokens, a text-only char/4 estimate that misses tool schemas,
   tool results, and images — badly undercounting long agent sessions.
   planner.Inputs gains ObservedInputTokens (the previous turn's billed
   input + cache-read + cache-creation tokens from the pin's usage
   writeback) and Decide prices tokens as max(estimate, observed): a
   monotonically growing agent prefix makes the prior turn's total a
   hard floor for this turn.

2. Planner EV shadow telemetry: the STAY/SWITCH verdict, reason,
   expected savings / eviction cost / threshold (USD), warmth
   assumption, and pinned from-model were only on OTel spans, which age
   out and don't join to session_pins or billed cost. Migration 0032
   adds nullable planner_* columns to model_router_request_telemetry
   (and recreates the production view), so every stay/switch
   counterfactual is replayable offline before policy changes ship.

No routing behavior changes: token grounding only feeds the existing EV
comparison, and the telemetry is write-only.

[aminsamir45]

• feat(planner): ground EV math in observed usage + persist EV verdict as shadow telemetry #577 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[greptile-apps]

T-Rex Logs

What T-Rex did

• Ran token-grounding tests, including the base run and the head run, and verified exit codes and behavior changes in observed savings and undercount calculations, plus a source check that shows how ObservedInputTokens are computed.
• Inspected planner telemetry schema changes, confirming the before state had zero planner_* columns and that an insert would fail due to the missing column, then confirmed the after state added the planner_* columns as nullable and that a planner row could be read back.
• Executed planner telemetry tests, noting that before there were no applicable tests and after running tests TestApplyPlannerTelemetry_SkippedLeavesFieldsNull, TestApplyPlannerTelemetry_StayRecordsEVBreakdown, and TestApplyPlannerTelemetry_SwitchPreservesFromModel all passed with exit code 0, and validated the STAY and SWITCH value mappings.

_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "feat(planner): ground EV math in observe..." | Re-trigger Greptile}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 69f31b0. Configure here.}

[greptile-apps]

T-Rex Logs

What T-Rex did

• The baseline token grounding was captured before the change from trex-artifacts/observed-token-grounding-01-before.txt.
• The after token grounding state was captured after the change from trex-artifacts/observed-token-grounding-02-after.txt, including updated EV, token counts, and EVComputed status.
• A proxy test verified Anthropic and OpenAI/OpenRouter token counts after the change, including observed_prompt_tokens and transcript counts.
• Evidence artifacts confirmed the attempted before run and the after/tooling blocker check, and noted that no contract mismatch occurred because the changed schema path was not executable in this environment.
• Planner telemetry tests ran, including the before/after telemetry writeback artifact and null semantics checks for multiple fields.

_{Ran code and verified through T-Rex}

_{Reviews (2): Last reviewed commit: "chore: retrigger CI (Test/Check Migratio..." | Re-trigger Greptile}