math: expose error sensitivity signal - #1

Summary

PR 3: make the feedback signal explicit. A run now exposes not just prediction + error, but the local sensitivity and the resulting correction pressure.

ŷ → E → f′ → E·f′ → feedback

• ActivationFunction.derivativeFromOutput(y) — sigmoid uses the y(1 − y) identity available straight off a forward pass (relu: y>0?1:0, identity: 1).
• errorSensitivity(activation, prediction, target) → { prediction, target, error, sensitivity, updateSignal }, where error = target − prediction and updateSignal = error · sensitivity.
• Example 03 rewritten to run the neural graph, then print E, f′, and E·f′ next to the evaluation and feedback action.
• Docs: explicit Input → … → Feedback Action diagram, the output-derivative identity, and worked values tied to example 02's prediction.
• Root examples script to run all examples in sequence.

Still no training/backprop — this only exposes the signal that would drive an update. This is the bridge: neural prediction error → derivative → update signal maps to harness run violation → diagnosis → update candidate.

Verification

• pnpm typecheck ✓
• pnpm build ✓
• pnpm test ✓ (40 tests; math now 16)
• Example 03 output: error E = 0.507856, f′ = 0.249938, update signal = 0.126933

Test plan

• math unit tests for derivativeFromOutput (sigmoid/relu/identity)
• math unit tests for errorSensitivity (signal fields + zero-error case)
• pnpm --filter @composable-model-graph/example-03-error-sensitivity-feedback start

Made with Cursor