Validator feedback names the fix, not just the rule (T0.4)

[GustavoSena]

Closes #27. Tier 0 of the prompt-building ladder (#29) — the independent, validator-side one.

Why

compose() feeds violation messages straight back as the next prompt, and there is exactly one retry before the deterministic fallback takes over. Several messages stated the rule without the value that satisfies it, leaving a 7B model to re-derive what it just got wrong. Runtime cost is zero — these strings are only built on rejection — and the validator still rejects, never rewrites (F2 §4): naming the fix is guidance for the next attempt, not mutation of the old one.

What changed

Every violation with a deterministic fix now ends by naming it:

Code	Now also says
I1	the tokens the user actually selected (symbol + address)
I2	divide the cap between strategies; the total must be ≤ cap
I3	the permitted count, both directions
I4	set "chainId": 8453
I5	the offered instruction — or "omit the slot slot" when the menu is empty
I5 (feeBps)	the scale that actually trips a model: out of 1e9, so 0.3% is 3000000
I7	the exact timestamp to use
I8	the known template ids
I10	the corrected token array and "move each virtualAmount with its token"
I11	what a valid amount looks like / that a leg must be positive
I12	the block to copy

Two are more than a value. I2 says to divide rather than repeat the budget — that is the "split by adding instead of dividing" failure the loop was built for. And an empty menu now reads as an answer ("nothing is offered here on this venue — omit the guards slot") instead of a truncated sentence; no guard has an encoder on this venue, so that is the honest fix.

Where no deterministic fix exists — which template suits the intent — the message states the rule and stops.

recommendation.ts's parse errors were audited too (they flow into the same feedback channel) and left alone: the prompt's OUTPUT_SCHEMA block already carries their fix, and duplicating it in the error text would just grow the retry prompt.

Verification

• 39/39 in validate.test.ts (13 new: one per code, each asserting the corrective literal — the sorted array, the timestamp, the address list — not just that a violation fired). Full SDK suite 144/144, typecheck clean.
• One new test pins a real subtlety: feeBps: 30 (0.3% in the 10000-scale a model knows) is a valid integer in [0, 1e9) and correctly does not fire — so the scale hint belongs to the out-of-range message, where it can actually help.

Not in this PR

The prompt-side Tier 0 work (#24, #25, #26) — separate PR, since it touches the prompt contract and bumps its version.

🤖 Generated with Claude Code