[c-RCP-AUDIT-03] Unify 5-file conditional-request split into one module (architecture target, judgment call on value)

[SoundMatt]

Background

RELAY's docs/RCP-ARCHITECTURE.md (canonical choice #4, "Conditional-request layer") names cpp-RCP's request.hpp and rust-RCP's request.rs as the reference shape: one request-kind enum plus shared sequencer-bank, priority-tier, and ledger state, all in a single module.

c-RCP instead keeps a 5-way file split, duplicating that shared state 5 times:

src/request_compound.c    src/request_chained.c    src/request_triggered.c
src/request_timed.c       src/request_cancel.c
include/rcp/request_compound.h   (+ matching headers for the other 4)

2353 total lines across these 10 files (2026-08-07 count).

Important: this is not a bug. This session independently re-verified c-RCP's Chained and Cancel implementations are byte-precise and TC18-conformant in their current split form — unlike go-RCP's equivalent layer, which has real wire-format defects (see go-RCP's own tracked redesign, a much larger and more urgent piece of work). This issue is purely about cross-repo architectural consistency, not fixing broken behavior.

Before starting: confirm this is worth doing at all

RCP-ARCHITECTURE.md itself calls this "the largest structural change in this effort — sequenced last, after the smaller items above are stable." Given:

• The current 5-file split is functionally correct today.
• A unification refactor of this size carries real regression risk for zero behavior change.
• There are other open items (issues [c-RCP-AUDIT-01] Wire TC18 Table 27's 16 unmapped error codes to real response paths #163, [c-RCP-AUDIT-02] Backfill TC18 citations on ~801 uncited requirements #164) with actual safety/traceability value and smaller blast radius.

The agent picking this up should explicitly raise with the user whether this is worth doing now, or at all, before writing any code — do not treat "RCP-ARCHITECTURE.md names it as the target" as sufficient authorization to start a large refactor unilaterally. This mirrors the standing project convention that large, invasive, non-bug-fixing changes get flagged for explicit human sign-off rather than started autonomously (see go-RCP's conditional-envelope redesign for the same posture on a much larger, bug-driven version of this problem).

If given the go-ahead, what the target shape looks like

1. One request-kind enum covering Compound, CompoundWait, Triggered, Chained, Timed, and the three cancellation forms (clear-all, clear-non-safestate, clear-single) — replacing the 5 separate per-file type systems.
2. Shared sequencer-bank, priority-tier, and ledger state owned once, not duplicated per request kind.
3. Reference the actual cpp-RCP (include/rcp/request.hpp) and rust-RCP (src/request.rs) source directly for the concrete shape — do not design from scratch or from this issue's description alone.
4. c-RCP's own request_compound.c/request_timed.c were separately named in RCP-ARCHITECTURE.md as reference shapes for wire-layer correctness (the field-level TC18 encoding, e.g. cmp_start_state/cmp_next_state/cmp_sequencer/cmp_exec_delay/cmp_repetitions) even while being flagged for restructuring at the module-organization level — preserve that correctness; this is an organizational refactor, not a wire-format change.
5. Requirement-id grouping (REQ-CMP-*/REQ-TRIG-*/REQ-CHAIN-*/REQ-TIMED-*/REQ-CANCEL-*) is already c-RCP's own convention and is itself the cross-repo target (canonical choice feat: v0.7.0 TLS transport — mutual TLS interface stub #7) — preserve these prefixes even as the underlying files merge.

Verification standard

• Full native build + full test suite + ASan/UBSan build (this repo has one, use it).
• Freshly built, CI-pinned cfusa: check 0 errors, trace --req-coverage 100 --sec-tested 100.
• Every existing test for Compound/CompoundWait/Triggered/Chained/Timed/Cancel must still pass unchanged in behavior — this is a pure structural refactor, so any test that needs to change its assertions (not just its location/imports) is a signal something broke.
• Mutation-test a sample of the moved logic (revert a small piece, confirm the now-missing behavior actually fails a test) to build confidence the refactor didn't silently drop coverage.

Not in scope for this issue

• Any wire-format or behavioral change — see issue [c-RCP-AUDIT-01] Wire TC18 Table 27's 16 unmapped error codes to real response paths #163 (error-code wiring) and issue [c-RCP-AUDIT-02] Backfill TC18 citations on ~801 uncited requirements #164 (citation backfill) for the actual open correctness/completeness gaps.
• Porting this shape to go-RCP — go-RCP's conditional-request layer has its own, much larger, already-tracked redesign (wire-format defects, not just organizational) that should not be conflated with this purely-organizational c-RCP issue.