feat: jagged pcs

[kunxian-xia]

Summary

This PR includes works in several separate PRs.

PR	Note
	compute q(b) from the input row major matrices
#32	run the jagged sumcheck $v = \sum_b q(b) * f(b)$ in a memory constrained device (note that both q and f have around 30 ~ 32 variables)
#39	evaluator for indicator function $g(z_1, z_2, z_3, z_4)$
	verifier for $f(z) = \sum_c \textrm{eq}(z_c, c) * g(z_r, z, t_c, t_{c+1})$
	jagged assist sumcheck
	stacked pcs (reshape q(b) as several smaller multilinear polynomials)

[hero78119]

@copilot please review this PR

[hero78119]

@copilot review this PR

[Copilot]

@copilot review this PR

Reviewed and fixed in commit b02d8c4:

1. Removed unused imports – use std::iter::once and IndexedParallelIterator were unused (the former only served the confusing scan trick). Replaced with FieldAlgebra which is needed for the safe initialization below.

2. Replaced unsafe { set_len } – changed to vec![E::BaseField::ZERO; total_size]. The parallel transpose loop unconditionally overwrites every slot, so this is zero-cost correctness: no behavioral change, no unsafe block.

3. Removed unreachable poly_heights.is_empty() guard – after the preceding validation (rmms non-empty, each matrix has non-zero columns), poly_heights is always non-empty at that point.

4. Simplified cumulative heights computation – the chain(once(&0)).scan(...) trick (appending a dummy 0 to coerce scan into emitting one extra value) is fragile and non-obvious. Replaced with a straightforward indexed loop:

   let mut cumulative_heights = vec![0usize; poly_heights.len() + 1];
   for (i, &h) in poly_heights.iter().enumerate() {
       cumulative_heights[i + 1] = cumulative_heights[i] + h;
   }

5. Added test_bitrev_transpose_layout – verifies the actual data layout after bit-reversal and column transposition against a known 4×2 input (col0=[1,5,3,7], col1=[2,6,4,8]). All 6 tests pass.

[hero78119]

@copilot rollback to 80d5537, then review with ".github/copilot-instructions.md" in base branch. just review, no edit/push. Author will fix it

[hero78119]

@copilot review this PR using .github/copilot-instructions.md in another branch feat/ci_ai_review

[Copilot]

Pull request overview

Adds an initial Jagged PCS adaptor to mpcs, including a commit-time packing of many column polynomials into a single “giga” MLE and a streaming jagged sumcheck prover intended for memory-constrained environments (per PR description and linked #32 context).

Changes:

• Expose a new mpcs::jagged module and re-export jagged commitment/sumcheck APIs from mpcs::lib.
• Implement jagged_commit (bit-reverse rows + transpose + concatenate) and jagged_sumcheck_prove (epoch-based M-table streaming + fallback to standard sumcheck).
• Add a Criterion benchmark for jagged sumcheck and introduce tracing-forest as a dev-dependency.

Reviewed changes

Copilot reviewed 4 out of 5 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
crates/mpcs/src/lib.rs	Exports the new jagged module and its public API types/functions.
crates/mpcs/src/jagged.rs	Implements jagged commit packing and streaming jagged sumcheck prover, plus unit tests.
crates/mpcs/benches/jagged_sumcheck.rs	Adds a Criterion benchmark for the jagged sumcheck prover across n=25..31.
crates/mpcs/Cargo.toml	Registers the new benchmark and adds tracing-forest dev-dependency.
Cargo.lock	Locks tracing-forest and related transitive dependencies.

[Copilot]

The module docs describe using an indicator function g(z_r, z_b, t[i-1], t[i]) to handle jagged segment boundaries, but the current implementation of f(b) throughout the prover is eq_row[row] * eq_col[col] (no dependence on t beyond (col,row) mapping). Either update the docs to match the implemented (uniform-height) assumption, or implement the g(...) weighting so variable-height segments are handled as described.

Suggested change

	//! The cumulative heights allow the verifier to succinctly evaluate the indicator function
	//! `g(z_r, z_b, t[i-1], t[i])` needed for the jagged sumcheck.
	//! In this adaptor, `t` is used for segment layout and for recovering the `(column, row)`
	//! coordinates of a packed evaluation. The prover logic then uses the standard product-form
	//! equality weights induced by those coordinates (effectively `eq_col[column] * eq_row[row]`).
	//! This implementation does **not** currently implement the more general boundary-weighting
	//! indicator `g(z_r, z_b, t[i-1], t[i])` described in the full jagged-sumcheck construction;
	//! accordingly, the documentation here is restricted to the uniform-height packing behavior
	//! implemented by this module.

[Copilot]

row/col from col_row_iter are used to index input.eq_row[row] and input.eq_col[col] without any validation. If eq_row.len() is smaller than the max polynomial height or eq_col.len() is smaller than num_polys, this will panic. Add an upfront check in jagged_sumcheck_prove (or a constructor for JaggedSumcheckInput) that eq_row.len() >= max_height and eq_col.len() >= num_polys, or change the API to return a Result with a clear error.

[Copilot]

Same issue as in build_m_table: input.eq_row[row] and input.eq_col[col] are indexed without checking that the precomputed tables are large enough for the (col,row) pairs implied by cumulative_heights. Without an upfront validation, this can panic for malformed inputs or when polynomials have larger heights than the provided eq_row table.

[Copilot]

test_jagged_sumcheck_n25 allocates and processes 2^25 (= 33,554,432) field elements (and additionally builds f_evals in final_evaluations_slow), which is likely to exceed typical CI time/memory budgets and make cargo test flaky/slow. Consider marking this test #[ignore], gating it behind a feature/env var, or moving it to a Criterion bench (keeping unit tests at small sizes that still cover the same code paths).

[Copilot]

Calling tracing_forest::init() inside a unit test performs global tracing subscriber initialization; this can conflict with other tests that also initialize tracing and can fail/panic depending on the tracing-forest behavior. Prefer a fallible/once-only init (e.g., via a Once guard or a try_init-style API) or remove tracing initialization from the test.

Suggested change

tracing_forest::init();