Test reset clears the shared-index built-roots ledger with the slots it describes

[gunbai-bot]

Cause

Not an order/thread flake — deterministic. shared_primary_precedence_index_is_byte_identical_to_cold fails run alone (--exact, --test-threads=1, no predecessor) on main since #12765 added PROCESS_RESOLVE_INDEX_BUILT, the per-thread ledger that turns "same roots built twice after eviction" into SharedIndexRebuiltAfterEviction. The thread-reuse hypothesis is falsified by that isolated run.

Chain (§6b): the ledger's contract is that a rebuild of roots the thread's slot held and lost to eviction is a redundant second parse — that refusal is correct and is kept unchanged. reset_process_shared_index_for_test is the one test-only operation that empties the slots to declare a fresh history (so the cold arm re-derives from disk). It cleared the slots but not the ledger that describes them, so after the oracle's cold arm + reset, the warm arm's rebuild of the same roots read as an eviction-rebuild. The earliest unjustified boundary is the reset: slots and ledger are one state and it reset half of it.

Fix

The reset clears the ledger with the slots. The refusal is not widened: production paths never call the reset, and a_shared_index_evicted_by_other_roots_refuses_to_rebuild (A, B, A without reset) still refuses.

Evidence (BuildBuddy, full paths)

• before: cli_run::compile_clean_via_index_verdict_equivalence::shared_primary_precedence_index_is_byte_identical_to_cold alone → FAILED (SharedIndexRebuiltAfterEviction at cli_run.rs:6668)
• after: same → ok; cli_run::closure_edge_demand_tests::a_shared_index_evicted_by_other_roots_refuses_to_rebuild → ok

🤖 Generated with Claude Code

[gunbai-bot]

Superseded by #12821, which deletes the built-roots ledger this PR would have cleared on test reset. Verified on BuildBuddy (--exact, --test-threads=1, no predecessor): cli_run::compile_clean_via_index_verdict_equivalence::shared_primary_precedence_index_is_byte_identical_to_cold is red on main 68ad067 (SharedIndexRebuiltAfterEviction at cli_run.rs:6668) and green on #12821 head ff25cc8. Branch kept.