[ocots]

Closes #356.

Summary

The code generated by @def and @init named CTModels, ExaModels and CTBase by symbol, so it only ran in a module that had bindings with these names: import OptimalControl or using OptimalControl: @def failed with UndefVarError: CTModels not defined. The module objects are now interpolated into the generated expressions, so they resolve where the macros are defined (same mechanism __wrap already used for CTBase.CTException, and the one applied to @Lie in CTLie.jl#43).

• CTModels, ExaModels and LinearAlgebra are direct dependencies (CTModels = "0.19, 0.20, 0.21", the range used across the ecosystem).
• The prefix machinery is replaced by three constants in src/defaults.jl: FUN = CTModels, EXA = ExaModels, ERR = CTBase.
• ext/CTParserExaModels.jl is folded into src/exa_linalg.jl (methods unchanged; still temporary, Align CTParser with CTBase 0.29 / CTModels 0.18 / ExaModels 0.12 #325).

Decisions (ct-conventions v0.10.0)

• init_prefix is removed too, beyond the literal list of the issue: @init had the identical defect (import OptimalControl + @init fails the same way) and nothing outside CTParser references it. Leaving it would only half-fix the bug.
• Aqua now sees the ExaModels glue (TST8: an extension escaped it). The deliberate piracy is whitelisted with treat_as_own on the five ExaModels node types, not on the whole module, so a new pirated method elsewhere still fails. The ambiguity check follows the ecosystem pattern (ambiguities=false in test_all, then Aqua.test_ambiguities(CTParser; exclude=...)), excluding the stdlib functions the port extends; those ambiguities are upstream's, ported verbatim, and leave with the file.
• Tests that can reproduce the bug (TST1, TST3): runtests.jl binds all three names in Main, so a flat test could not fail before the fix. test/test_hygiene.jl uses nested modules where the packages are loaded but the names are unbound (what import OptimalControl leaves), plus a module binding CTModels = 42, ExaModels = "nope", CTBase = nothing. It covers @def (:fun and :exa), @init, and the typed errors raised from generated code.
• test_prefix.jl is deleted, test_prefix_bis.jl becomes test_backends.jl with a contract on FUN/EXA/ERR and an isdefined guard against the removed API coming back.

Breaking

The prefix API is removed (prefix_fun(!), prefix_exa(!), e_prefix(!), init_prefix(!), the __default_*prefix* functions and the PREFIX_* constants). Documented in CHANGELOG.md and BREAKING.md with a migration block (CHG3). The only external caller found is the archived ProjetLong26/CTTransfo. ExaModels is now loaded with CTParser, including for users of the functional backend only (it is already a dependency of OptimalControl).

Test plan

• Pkg.test("CTParser"): 15/15 files, 2627 pass, 9 broken (the expected GPU @test_skip), 0 fail
• test_hygiene.jl fails on main's generated code (module Bare raises UndefVarError), passes here
• CI (including the GPU runners)
• Downstream: OptimalControl no longer needs to re-export CTModels/ExaModels/CTBase for the generated code (separate PR there)

Conventions: ct-conventions v0.10.0, ct-agents v0.4.0.

Generated with Devin

[ocots]

Review

Reviewed at c5049f0 against ct-conventions v0.10.0 / ct-agents v0.4.0.

Verdict: good to merge once the remaining test-plan items are done. The fix is right: it uses the mechanism __wrap already used for CTBase.CTException, and it is applied everywhere. I grepped src/ and found no expression that still names CTModels, ExaModels or CTBase by symbol; every pref / e_pref is now FUN / EXA / ERR. Removing init_prefix too was the right call: leaving it would have kept the bug in @init.

What I checked

• Generated code. def_fun, def_exa (builder, dyn_check, getter), __throw, _bound_length_check, _unknown_scheme_error, all p_*_fun! / p_*_exa! and init_fun / __gen_temporal_value interpolate the module objects. The macros still esc the whole expression, which is harmless now that the modules are objects and not names to resolve.
• Tests. test_hygiene.jl is the test that was missing. It can fail (the Bare module reproduces import OptimalControl), and Shadowed covers the worse case where the caller binds those names to something else. The fixture modules sit at the top level of the test module (TST3).
• CTModels lower bound. Everything the generated code calls (PreModel, definition!, time_dependence!, build(...; build_examodel=), time_name, build_initial_guess, validate_initial_guess, get_build_examodel) already exists at v0.19.0-beta, so 0.19, 0.20, 0.21 looks sound. CI only resolves the latest version, though, so the lower bound itself is not tested.
• Aqua. treat_as_own on the five node types is tighter than whitelisting the whole ExaModels module, and it also covers the ExaModels._hdrpass_val method.
• CI. CPU passes on 1.10 / 1.13 × 3 OS, and test-gpu-kkt passes. test-gpu-main / test-gpu-occidata were skipped, so the GPU box in the test plan is still open.

Minor points (none blocking)

1. Stale docstring: src/initial_guess.jl:100 (__gen_spec_value) still says - `pref::Symbol`: backend module prefix. It should read pref::Module like the __gen_temporal_value docstring just above it.
2. The Aqua ambiguity exclusions are broad. exclude=[Base.:*, Base.:+, Base.:-, Base.convert, Base.zero, ...] turns off the ambiguity check for those functions entirely, not just for the methods in exa_linalg.jl. That is fine while the port exists, but these exclusions must be removed together with the file. Could you add that explicitly to Align CTParser with CTBase 0.29 / CTModels 0.18 / ExaModels 0.12 #325 ("delete src/exa_linalg.jl and restore ambiguities=true / drop treat_as_own in test/test_aqua.jl") so that someone cleaning up later does not miss it?
3. Optional nit: pref = FUN / e_pref = ERR are now plain aliases of constants and could be inlined ($FUN.PreModel()). Keeping them keeps the diff small, so it is up to you.
4. Optional, test coverage: Shadowed does not exercise __throw (the ParsingError raised at expansion time by @init). Bare.broken_init covers it with the names unbound, but not with the names bound to junk values. A one-line Shadowed.broken_init would make the two modules cover the same cases.

Release note

This PR is breaking, so the next release must be a breaking bump (a minor bump in 0.x). Project.toml still reads 0.8.17-beta on purpose (it gets bumped in the release PR), while the latest tag is v0.9.5-beta, so the release PR should take care to pick the right version. The downstream OptimalControl PR (dropping the re-exports needed by the generated code) can be opened in parallel.

🤖 Generated with Claude Code