[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Summary

Fixes #5200: discrete equational affects were tearing for parent differentials instead of re-solving algebraic variables. Pin a differential only when the parent algebraic/observed equations that contain it can be bipartite-matched to distinct free (non-user-assigned) algebraic unknowns. Never pin dummy-derivative variables from index reduction, nor other differentials that appear alongside them in differentiated algebraic constraints (those equations must stay matched to free unknowns — otherwise root InterfaceI pendulum complete throws UnsolvableCallbackError). Never pin a differential that appears in the user's affect outside Pre. Array differentials use AffectPinOrigin stand-ins so setsym never receives Pre of an indexed/array variable. OOP SVector support is split to #5246. Events.md is unchanged.

Behaviour change for Chris: in a DAE whose differentials appear in algebraic equations, a discrete-only affect now keeps those differentials fixed and re-solves the algebraics. On master the same affect could move the differential. Existing models can get different trajectories. This is what #5200 asks for, but it is a silent semantics change while Events.md rule 1 still says every unknown is modifiable unless x ~ Pre(x).

Fail before / pass after

Master callbacks only vs this PR:

Case	Master	This PR
#5200 scalar on~0	v 0.716→0.316	v stays 0.716; λ→−0.358
Array u[1:2] + alg + on~0	compiles; u[1]→~0	compiles; u unchanged; λ≈u[1]
Probe A: two algs share one free absorber, assign λ	Success; after event x=2, λ=3, μ=−1	same (no pin; no UnsolvableCallbackError)
Residual 0~v+0.5*Pre(v)	0.5→−0.25	same
Inelastic collision	v→0	same
Docs inverse y~Pre(y)+1	[1,1]→[0.5,2]	same
Pendulum Symbolic affects are compiled in complete (root InterfaceI)	passes	passes (no longer pins x/xˍt)

CI fixes on this head (bb771781c1)

1. Never pin dummy derivatives / index-reduced constraint variables — schedule.dummy_sub and algebraics that mention those stand-ins are banned from pinning so the affect matching stays square.
2. Trial-pin fallback — if substituting Pre would leave an algebraic with no free unknown, skip the pin instead of throwing UnsolvableCallbackError.
3. Symbolic sol indexing — Discrete event on a system with algebraic equations changes a differential state instead of the algebraic variable #5200 / collision / inverse regressions use sol[x], sol[λ], sol[μ] (unknown order is not stable across simplification paths).

Verification (stacked private depot; one Julia process at a time)

• Top-level GROUP=InterfaceI +1.13.1: InterfaceI | 1588 Pass, 3 Broken (~33m) — EXIT 0.
• Top-level GROUP=InterfaceI +1.12: InterfaceI | 1589 Pass, 3 Broken (~45m) — EXIT 0.
• lib/ModelingToolkitBase GROUP=InterfaceI +1.13.1: InterfaceI | 1805 Pass, 5 Broken (~29m) — EXIT 0.
• lib/ModelingToolkitBase GROUP=InterfaceI +1.12: InterfaceI | 1805 Pass, 5 Broken (~39m) — EXIT 0.

Risk assessment

• Risk: medium — intentional Discrete event on a system with algebraic equations changes a differential state instead of the algebraic variable #5200 pin changes DAE+discrete-only affect trajectories; Chris should decide if breaking.
• Blast radius: AffectSystem pin construction only; no public API signature change; Events.md unchanged.
• Evidence: master-vs-PR probes; pendulum root InterfaceI green; symbolic indexing regressions.
• Merge: needs review.

Fixes #5200 (algebraic resolve). OOP: #5246.

Please ignore this draft until reviewed by @ChrisRackauckas.

Harness: Cursor Agent CLI 2026.10.01-14929f9, model auto; transcript <job>/log.txt.

Review findings addressed

Review of 80efebf (codex-cli gpt-6-astra, verdict CHANGES), fixed in the follow-up commit:

1. All-or-nothing feasibility disabled pins in unrelated components — fixed. Pin selection now lives in affect_pin_substitutions. It builds a maximum matching of the affect equations (user, algebraic and observed) to the affect unknowns, using non-differential unknowns first and untouched differentials only where no other unknown can absorb an equation (augmenting paths over maximal_matching). Candidates left unmatched are pinned, so each independent part of the system is decided on its own. The old global can_pin_candidates/pin_ok checks are gone; the matching guarantees every equation keeps an unpinned unknown. The reviewer's composed model (#5200 plus D(z) ~ 1e-16, z*y ~ p, affect [on ~ 0, y ~ Pre(y)+1]) now keeps v at 0.7159745833 (head: → −1.2840254167) while z → 0.5, y → 2. Regression testset: "Issue#5200: differentials are pinned per independent part of the system".
2. An indexed element update unpinned its siblings — fixed. Symbol sets are now SymbolFootprints: u[1] and u[2] with constant indices do not overlap, while a whole-array reference or a non-constant index still overlaps every element. This applies to the touched set, the dummy-derivative/index-reduced bans and the candidate filter. Array references that are not element-wise (whole array or symbolic index) in algebraic/observed equations are never pinned. Regression testset: "Issue#5200: assigning one array element keeps the other elements pinned" (head: u[2] 0.7159745833 → 0.3159745833; fixed: unchanged). It uses u[1] ~ 2.0 rather than the reviewer's u[1] ~ Pre(u[1]) + 1: under ModelingToolkitBase alone, an implicit affect containing Pre(u[1]) fails with Invalid symbol Pre((u(t))[2]) for setsym on master 2b821a5, on this PR's previous head and with this fix alike. That is a separate, pre-existing bug. Under ModelingToolkit the reviewer's exact Pre form now keeps u[2] (0.7159745833 → 0.7159745833, u[1] +1).
3. Undocumented change of the public callback policy — docs/src/basics/Events.md now states the pinning as rule 3 and gives the #5200 example with the opt-out (λ ~ Pre(λ) holds the algebraic variable and lets v change; checked locally: v 0.716 → −1.284, λ stays 0.642). No version bump here (the repo bumps versions in separate release PRs); see Breaking change below.

Verification (Julia 1.13.0, JULIA_NUM_THREADS=4)

Fail before (previous head 80efebf) / pass after, for the #5200 testsets extracted from symbolic_events.jl and run under both packages:

Mode	Previous head	This commit
ModelingToolkitBase only	46 pass, 3 fail	49 pass
ModelingToolkit	46 pass, 3 fail	49 pass

The three failures before the fix were any(isequal(v), parameters(aff)) (evaluated with [p] only), sol[v] -1.2840254166877405 ≈ 0.7159745833122596, and sol[u[2]] jumping from 0.716 to 0.316.

Full suites on this commit:

• Root GROUP=InterfaceI julia +release --project -e 'using Pkg; Pkg.test()': InterfaceI | 1640 Pass, 3 Broken, 1643 Total (26m35s), ModelingToolkit tests passed. The 3 Broken are pre-existing @test_brokens. This group also runs symbolic_events.jl under ModelingToolkit, including the index-reduced pendulum complete case.
• ModelingToolkitBase-only run of the full lib/ModelingToolkitBase/test/symbolic_events.jl: Symbolic Event Test (ModelingToolkitBase only) | 458 Pass, 2 Broken, 460 Total (7m45s). The 2 Broken are the existing @test_brokens in the "Bump" testset.

Formatting: Runic --check is clean on the changed Julia files. typos is clean on all changed files.

Not run locally: GROUP=QA (the only import change adds construct_augmenting_path!, which BipartiteGraphs exports), the docs build (Events.md gains prose and a non-executed julia block), the full lib/ModelingToolkitBase GROUP=InterfaceI, GPU and downstream groups.

Breaking change

For an equational (symbolic) affect, a differential variable that the affect does not reference outside Pre is now held at its pre-event value whenever the algebraic equations can be re-solved through algebraic unknowns or discrete parameters instead. Before this PR, compiling the affect system could choose to move such a differential state instead (the #5200 behaviour). This applies to both discrete and continuous callbacks. Models that relied on the old choice get different post-event states. They can restore it by adding alg ~ Pre(alg) for the algebraic variable that should stay fixed. Events.md rule 1 ("all unknowns are treated as modifiable") was the documented contract and is now refined by rule 3. Under strict SemVer this is a default-behaviour change of public API and needs a major bump (ModelingToolkitBase 1.x → 2.0, and correspondingly ModelingToolkit), unless the maintainer classifies the old behaviour as the bug that #5200 describes and ships it as a fix. This PR does not bump any version.

---

🤖 Posted by an AI agent — harness: Claude Code 2.1.301 · model: claude-opus-5-5
Conversation: local session ID 56988e15-46f1-419c-a0f8-7f1e94a5aa67 (headless, no URL)

Addendum (after push): lib/ModelingToolkitBase GROUP=InterfaceI on e87832c gives InterfaceI | 1821 Pass, 1 Error, 5 Broken (29m48s), with Symbolic Event Test | 458 Pass, 2 Broken. The single error is Parsing Test (variable_parsing.jl:66, Variable x must have default of matching size). It comes from Symbolics v7.46.0 rejecting x = [1, 2] on a scalar variable. Master already updated that test in 06b51a13be (#5281), after this branch's merge base, so it is base-branch drift unrelated to this change. Merging master into the branch would pick up that fix.

---

🤖 Posted by an AI agent — harness: Claude Code 2.1.301 · model: claude-opus-5-5
Conversation: local session ID 56988e15-46f1-419c-a0f8-7f1e94a5aa67 (headless, no URL)

Risk assessment (independent review, Grok Build grok-4.7, reviewed head e87832c)

• Risk: medium
• Blast radius: Symbolic discrete and continuous affects share AffectSystem in ModelingToolkitBase. A differential that the affect does not reference outside Pre is now held when an algebraic unknown or discrete parameter can absorb the equation, decided per independent part of the system. Existing DAE models that used equational callbacks can get different post-event states. No exported name is added or removed. Versions stay ModelingToolkit 11.45.4 and ModelingToolkitBase 1.77.4. BipartiteGraphs is already a dependency; the new uses are its existing maximal_matching and construct_augmenting_path!.
• Evidence: Julia 1.13.2, JULIA_NUM_THREADS=4, local develop of this head. ModelingToolkit mode: composed model keeps v at 0.7159745833122596 while z goes 1.0→0.5 and y goes 1.0→2.0; u[1] ~ Pre(u[1])+1 sends u[1] 0.6065306597126487→1.6065306597126487 and leaves u[2] at 0.7159745833122577; scalar Discrete event on a system with algebraic equations changes a differential state instead of the algebraic variable #5200 leaves v at 0.715974583312258. Opt-out λ ~ Pre(λ) sends v 0.715974583312258→-1.2840254166877418 and the post-event sample of λ is 0.6420127083438709, the pre-event consistent value. Vector algebraics λ = u - on keep both u elements and recompute both λ elements (at t=0.49, λ ≈ u-1; at t=0.5, λ ≈ u). Pendulum affect has 5 equations and FBDF returns Success. Same probes under ModelingToolkitBase alone: composed and scalar pins hold; u[1] ~ 2.0 keeps u[2]; Pre(u[1]) still hits the old setsym error. CI at review end: 45 checks pending, 0 failed, on the recorded head. Runs: https://github.com/SciML/ModelingToolkit.jl/actions/runs/38932552312 https://github.com/SciML/ModelingToolkit.jl/actions/runs/38932551242 https://github.com/SciML/ModelingToolkit.jl/actions/runs/38932551849 https://github.com/SciML/ModelingToolkit.jl/actions/runs/38932551986
• Independent review: Grok Build grok-4.7 rated it medium, high confidence: the two previous successful-but-wrong jumps are gone on local solves, and the documented opt-out moves v while holding λ.
• Merge: needs human review — the behavior change is what issue Discrete event on a system with algebraic equations changes a differential state instead of the algebraic variable #5200 asks for and the PR flags it, but ModelingToolkitBase is 1.x and this PR does not bump the version. A maintainer has to accept that release classification. Pending CI is not a reason to block the verdict.

---

🤖 Posted by an AI agent — harness: Grok Build · model: grok-4.7 (independent reviewer; appended by Claude Code 2.1.301 · claude-opus-5-5)
Conversation: https://claude.ai/code/session_01VZjwf7Ux9YtTzwHGHMqnsq

[ChrisRackauckas]

🤖 Automated comment from an AI agent running as @ChrisRackauckas — not written or reviewed by Chris.

Merged the default branch into this PR to resolve conflicts (head 138fb1c). Re-review the resolution before merging.

Conflict resolution report

STATUS: RESOLVED

Conflicting files

• lib/ModelingToolkitBase/test/symbolic_events.jl: both sides added new testsets at the end of the file, after the same final testset.
  • PR side: "Issue#5200: discrete affect re-solves algebraic vars, keeps differentials", "Equational affects: residuals and documented inverse keep differentials free", "Two algebraics sharing a differential stay free when only one is assigned", "Issue#5200 array differential with algebraic and discrete-only affect".
  • Master side: "Symbolic array index in affect (issue [bot-unreviewed] Indexing an array parameter with a discrete Int in a symbolic affect fails in mtkcompile (StableIndex typeassert) #5078)", the two "Issue#5077" testsets, and "User callback with discrete events".
  • Resolution: kept both. The PR's testsets come first, then master's, and each block keeps its own closing end. Nothing was edited inside either block. I formatted the file with Runic (julia +1.12 --project=@runic -m Runic -i), which changed nothing.

All other files (including lib/ModelingToolkitBase/src/systems/callbacks.jl, which both sides touched) merged automatically.

Tests (Julia 1.12.7, run one at a time)

Setup: julia --project -e 'using Pkg; Pkg.instantiate()' and julia --project=lib/ModelingToolkitBase -e 'using Pkg; Pkg.instantiate()'.

1. ModelingToolkitBase on its own:
   JULIA_LOAD_PATH="@:@stdlib:$HOME/.julia/environments/testenv" julia --project=lib/ModelingToolkitBase -e 'using TestEnv; TestEnv.activate(); using Test; @testset "symbolic_events" begin include("lib/ModelingToolkitBase/test/symbolic_events.jl") end'
   Result: symbolic_events | 445 Pass, 2 Broken, 447 Total. 0 fail, 0 error.
2. Through top-level ModelingToolkit, the way test/group_interfacei.jl runs it: root project, TestEnv, using ModelingToolkit; import ModelingToolkitBase; include(joinpath(pkgdir(ModelingToolkitBase), "test", "symbolic_events.jl")) inside @testset. pkgdir(ModelingToolkitBase) points at the local lib/ModelingToolkitBase.
   Result: symbolic_events | 464 Pass, 2 Broken, 466 Total. 0 fail, 0 error. The 19 extra tests are the @isdefined(ModelingToolkit) branch of master's [bot-unreviewed] Indexing an array parameter with a discrete Int in a symbolic affect fails in mtkcompile (StableIndex typeassert) #5078 testset.

The 2 Broken results were already marked @test_broken in the "Bump" testset before this merge. They are not new.

Logs: scratch/mtkbase_symbolic_events.log, scratch/mtk_symbolic_events.log.

Not run: the full GROUP=InterfaceI suites and QA. Only the test file containing the PR's new tests was run.

---

🤖 Posted by an AI agent — harness: Devin CLI 3000.11.3 (merge), Claude Code 2.1.285 (posting) · model: fusion-claude-opus-5-5-high-sidekick-swe-2-medium
Conversation: local Claude session 3c311569-dda6-4687-bc9f-65febb212c33 (fleet master)