predict: add frozen-grid inventory impact

[sdelo]

Summary

• Live mint and close now compute an inventory-impact charge from a frozen 100-bucket capital coordinate (K = average(top 5 bucket maxima) − E) and a capped convex potential. The configured rate is zero, so the path is live and the charge is still zero.
• initialize_inventory_grid / refresh_inventory_grid are authenticated keeper entry points. Callers submit 99 interior strike/forward ratios; the contract materializes the absolute ladder against its own forward and checks each bucket mass in [0.99%, 1.01%].
• Refresh rebuilds from an inline 2048-cell geometric array rather than walking payout-tree dynamic fields, so a full book stays under the object-cache ceiling.

Why

Inventory impact prices the extra concentrated capital a mint or close leaves on the book. The earlier L-based design with rebates was withdrawn; this replacement freezes a 1%-mass grid at initialize/refresh and charges only when centered capital rises. The rate stays at zero so the machinery can land and be gas-tested without turning the fee on. Refresh used to walk the payout tree and hit the object-cache wall around 1,000 nodes; the inline cell array is the path that completes a full book.

Linear / Context

• N/A — exempt (no DBU ticket).
• Supersedes the withdrawn L-based inventory-impact work on predict: add path-independent inventory impact #1226.

Key decisions

• Default-off. inventory_impact_max_rate = 0. This is not a turn-on. Calibration of B_K / r_K is out of scope.
• Frozen grid, not a live tree walk for the charge coordinate. The 100-bucket 1%-mass ladder is frozen at initialize/refresh. Incremental mint/close apply quantity into the cells and re-cut buckets from that snapshot.
• Ratio boundaries, not absolute strikes. The caller submits 99 interior ratios so the contract can re-anchor the ladder to the pricer's forward and reject a mass-check miss.
• Inline cell array instead of payout-tree refresh. A 16 KB inventory_cells vector lives on the market object so refresh reads zero payout-tree children. Incremental E uses the same snapped cell-span probability as apply, so refresh onto the same snapshot does not jump K.
• No rebates, no escrow. The charge is ordinary expiry cash. E at close is re-derived from the current frozen snapshot, not stored at mint.

Scope / Descoped

• In scope: Move source, unit and flow tests, sessions max_cost wiring, harness keeper / PTB / campaign traces, simulation event decode for the new charge field.
• Out of scope: public design docs, notebooks, figure assets, predeploy evidence and register updates (left uncommitted).
• Deferred: nonzero rate, B_K / r_K calibration, flat-vs-convex policy, liquidation inventory-charge policy, late-life cell resolution, nonzero-rate quote gas.

Test plan

• sui move test --path packages/predict --gas-limit 100000000000 — 513 passed, including inventory-grid, inventory-cells, and inventory-impact flow tests.
• sui move test --path packages/sessions --gas-limit 100000000000 — 24 passed (redeem_live now takes max_cost).
• sui move build --path packages/predict --warnings-are-errors — clean.
• Localnet capacity-tree with KEEPER_INVENTORY_GRID=1 — grid init ~40M compute; refresh at a full book peaked at 626M compute (13% of the 5,000M cap) and did not abort on the object-cache wall.

Risk / Rollout

• Public markets stay at rate zero. A misconfigured nonzero rate would start charging live mint/close; protocol config still owns the bound.
• Refresh can abort inventory_grid:2 if a submitted ratio ladder fails the mass check after an SVI roll; the next keeper tick retries.
• Sessions redeem_live gained a max_cost argument. Callers of that entry must update with this package upgrade.
• Rollback is a package revert; no on-chain inventory-grid state exists on currently deployed markets until a keeper initializes one.

Made with Cursor