1. Most turns: finding out that the checker sees the literal tier as a view. The tier pass (crates/nikaia/src/text_tiers.rs mark) rewrites the parsed types before the checker runs, so a_map_read_kept returned early on read.is_a_view(). I spent about 8 turns reading check_core_leaves (the lending path), check_values, check_items and views before I grepped lasting in the Rust code. Second: I found late that return m[k] failed the same way, which cost one more build cycle.
2. Lookups the brief could have handed me:
   • That the tier pass runs at parse time and turns a literal-tier String into is_view && lasting, so the checker’s Ty has view: true and cannot tell a lasting view from a borrowed one.
   • That map_read_sites is set in check_core_expr.nika:88, with the raw read type at hand there.
   • That the bracket form m[k] fails too. The brief said only m.get(k).
3. Checks run:
   • Lowering probes with the old binary, which were useful.
   • Three build plus lower-std cycles: two for the fix, plus one fixpoint round because fold_bound.rs changed.
   • nikaia run on 7 probes, then the gate (clippy, guard set, tests/compiler at both settings), then pre-push-check.

   Wasted: the first fix was method-only, so one whole cycle went before I extended it to m[k]. Running the m[k] probe at the start would have saved it. The numbers probe (p7) answered nothing the brief needed.

4. One change: builder-facts should say: “the text tier pass runs before the checker. The literal tier is is_view + lasting on the parsed type, so a checker Ty sees it as a view.” A brief about a text-tier defect should also list the bracket form m[k] among the cases to probe. One probe matrix with every spelling (get, [], let, typed let, argument) written before the first build would have saved a cycle.

This site is open source. Improve this page.