Order book: take the PDA bump search out of the depth-cost test

[mikemaccana]

The flake

deepest_path_adds_little_compute_to_insert_fill_and_cancel compares an insert and a fill at depth 64 with the same instruction on a shallow book, and allows 15,000 extra compute units. In one CI run on #171 the Anchor v1 insert measured 15,228 extra. That program was identical to main's, and the re-run passed.

Cause

Both instructions create an order account with a bare bump, so Anchor searches for the bump at runtime and pays for every rejected seed. The market's address comes from generated mints, and tests running in parallel draw from shared key generation. So the search length changes from run to run, and with it the measured difference.

I reproduced this locally: the same code measured 37,306 and 38,806 units at depth 64 in different runs, exactly one 1,500-unit search step apart.

Fix

The test computes each order PDA's bump itself and subtracts the cost of the search:

Version	Cost per rejected bump	Why
Anchor v1	1,500 units	sol_try_find_program_address
Anchor v2	334 units	sol_sha256 plus the curve check, measured

With the search taken out, the extra cost of depth 64 is the same on every run:

Version	Insert	Fill
Anchor v1	about 10,650	about 10,660
Anchor v2	10,538	10,730

The 15,000 limit is unchanged.

Checks

• I built both programs with cargo build-sbf (Agave 3.1.8).
• Both full suites (34 tests each) passed five times in a row, with at most 31 units of jitter.
• cargo fmt --check is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JGEoAUjMm7Evv69k46eNcn

---

Generated by Claude Code