rtl-p4-implement-policy

Phase 4 Implementation Policy

10-Wave Pipeline Definition

Wave 0:  Prepare      — **Wave 0: Preparation + Test Plan**
         - Step 0a: Module enumeration, directory creation, TODO list (existing)
         - Step 0b: Test plan generation per module via test-plan-writer agent
           - Gate: sim/{module}/{module}_test_plan.md exists for every module
           - Failure: retry once, then proceed with WARNING (Wave 6a must generate missing plans)
Wave 1:  Write All    — One rtl-coder per module, all parallel
Wave 2:  Lint All     — One lint-checker per module, all parallel, collect results
Wave 3:  Fix Lint     — ONLY FAIL modules, max 3 rounds, re-lint only fixes
Wave 3.5: Synth Gate  — synthesizability-gate per lint-clean module, parallel (HARD GATE:
         zero inferred latches / incomplete assignments / non-synth constructs, AND
         DC-script-emittable; max 2 fix rounds, then rtl-architect escalation)
Wave 4:  Code Review  — rtl-critic per lint-clean module, parallel
Wave 5:  Bugfix       — ONLY REVIEW_FAIL modules, max 3 review→fix iterations
Wave 6a: Tier 1 Smoke — testbench-dev + eda-runner per module, parallel
Wave 6b: Tier 2 Unit  — Ref model comparison, REQ-U-* tracing, FSM >= 50%, line >= 60%, covergroups >= 1, codec conformance PASS/N/A
Wave 7:  Module CDC   — cdc-checker per multi-domain module, parallel
Wave 8:  Module Proto — protocol-checker per bus-interface module, parallel
Wave 9:  Refactoring  — rtl-p4s-refactor for flagged modules, selective
Wave 10: Integration  — smoke test + spec compliance + Stream B + Phase 4 gate

Key principles:

• "Lint all at once, fix only failures, re-lint only fixes"
• "Code review before testing — catch design bugs early"
• "Module-level CDC/protocol before Phase 5 — catch hazards early"
• "Stream B synthesis smoke test — catch latch inference and unmappable constructs before Phase 5"
• Modules that pass early waves start later waves immediately

Wave Overlap Rules

Owned by the wave sequencer: see "Wave Overlap Rules" in the p4-implement-orchestrator agent prompt (modules progress through waves independently per that schedule).

Per-Module State Tracker Schema

{
  "module": "{module}",
  "wave_1_write": "DONE",
  "wave_2_lint": "PASS|FAIL",
  "wave_3_fix": "PASS|SKIP",
  "wave_4_review": "REVIEW_PASS|REVIEW_FAIL",
  "wave_5_bugfix": "PASS|SKIP",
  "wave_6a_tier1_smoke": "PASS|FAIL",
  "wave_6b_tier2_unit": "PASS|FAIL",
  "wave_7_cdc": "CDC_PASS|CDC_FAIL|SKIP",
  "wave_8_protocol": "PROTOCOL_PASS|PROTOCOL_FAIL|SKIP",
  "wave_9_refactor": "DONE|SKIP",
  "wave_10_gate": "PASS|FAIL"
}

Coding Convention Summary (Core Overrides)

All RTL produced in Phase 4 MUST follow:

• Port prefix: inputs i_, outputs o_, bidirectional io_ (NOT suffix _i/_o)
• Clock: clk (single domain) or {domain}_clk (e.g., sys_clk) — NOT clk_i, clk_sys
• Reset: rst_n (single domain) or {domain}_rst_n (e.g., sys_rst_n) — NOT rst_ni
• Clock/reset ports do NOT need i_ prefix
• logic only — reg and wire keywords FORBIDDEN
• always_ff for sequential, always_comb for combinational — no bare always
• typedef enum logic [N:0] for FSM states, typedef struct packed for grouped signals
• Instance prefix: u_, generate prefix: gen_
• Parameters: UPPER_SNAKE_CASE, localparam: L_ prefix, types: snake_case_t
• ANSI port style, one module per file

Memory Wrapper Rules

Storage elements specified as "SRAM wrapper" in Phase 3 μArch docs MUST use the standardized wrappers in rtl/common/ (parameterized behavioral models for simulation, foundry macros for synthesis):

Wrapper	File	Use
sram_sp	rtl/common/sram_sp.sv	Single-port: 1 R/W port, single clock
sram_tp	rtl/common/sram_tp.sv	Two-port: 1W + 1R, single clock (clk)
sram_dp	rtl/common/sram_dp.sv	Dual-port: 1W + 1R, dual clock (wclk/rclk)

• Parameters: DEPTH, WIDTH; derived ADDR_W = $clog2(DEPTH) (localparam inside wrapper)
• Instance naming: u_mem_{purpose} (e.g., u_mem_coeff, u_mem_line_buf)
• Full interface spec (per-wrapper port lists, storage selection thresholds, synthesis blackbox/macro-replacement strategy) is OWNED by .claude/rules/rtl-coding-conventions.md ("Storage Selection" section, deployed by rat-init-project) — follow it verbatim
• Wave 1 responsibility: rtl-coder creates rtl/common/sram_*.sv wrappers if not already present, then instantiates them in modules per μArch spec

Combinational Chain Depth Heuristic

Sequential dependency loops in always_comb (e.g., error diffusion, SAD accumulation, carry chains) create deep combinational paths that risk timing closure.

• Threshold: > 4 sequential-dependency iterations in a single always_comb block is a timing risk
• If uArch budgets extra latency: split into sub-stages with pipeline registers at N/2
• If uArch requires single-cycle: do NOT split — escalate to rtl-architect/timing-advisor for review
• Exceptions: FPGA carry chains, balanced adder trees, and low-frequency clock domains may tolerate deeper chains — document the justification

Pattern: for (s = 0; s < N; s++) where iteration s depends on iteration s-1 → split at N/2 with registered intermediate results (when latency budget allows)

Examples where split was beneficial (ASIC):

• Error diffusion: 4-sample chain × 2 sub-blocks → split into 2+2
• SAD accumulation: 8 abs-diff + sum → register after 4
• Carry chain: 16-bit ripple → register at bit 8 (ASIC only — FPGA carry chains are an exception)

Code Review Focus Areas (Wave 4)

Per-module review by rtl-critic:

1. uarch compliance: does RTL match docs/phase-3-uarch/{module}.md?
2. Interface compliance: do ports match io_definition.json?
3. FSM completeness: all states from uarch spec present?
4. Pipeline correctness: stage count, latency, throughput match uarch?
5. Coding style: naming conventions, parameterization, comments
6. Logical correctness: off-by-one, sign extension, width mismatches
7. Power: unnecessary toggling, missing clock gating opportunities

Classification: REVIEW_PASS (0 critical/major findings) or REVIEW_FAIL.

Structural Verification in Code Review (Wave 4)

Wave 4 code review MUST include automated structural checks (performed by rtl-critic):

1. FSM state completeness: all uarch spec states present in RTL
2. Pipeline depth: RTL stages match uarch spec ±1
3. Port mapping: all uarch spec ports present with correct direction/width
4. Timing contract: pipeline supports specified latency (if documented)

Structural findings are classified:

• CRITICAL: missing FSM state, missing port → triggers Wave 5 bugfix
• MAJOR: pipeline depth mismatch, timing violation → triggers Wave 5 bugfix
• MINOR: naming inconsistency → advisory only

REVIEW_PASS requires zero CRITICAL structural findings.

Code Review Iteration Protocol (Wave 5)

• Round 1: Full review (all focus areas)
• Round 2: Targeted re-review (only previously failed focus areas)
• Round 3: Final check (must pass or escalate)

CDC Check Scope (Phase 4 vs Phase 5)

• Phase 4 (Wave 7): Module-level CDC — within each module boundary
• Phase 5 (rtl-p5s-cdc-verify): System-level CDC — across module boundaries, top-level analysis
• Phase 4 catches module-internal hazards early; Phase 5 catches inter-module hazards

Protocol Check Scope (Phase 4 vs Phase 5)

• Phase 4 (Wave 8): Module-level protocol — each module's bus interfaces in isolation
  • Timing contract assertion checks for interfaces with timing specs in uarch
  • valid/ready backpressure exercise
  • Multi-beat transfer protocol verification
• Phase 5 (rtl-p5s-protocol-verify): System-level protocol — end-to-end transaction flow
• Phase 4 catches per-interface violations + timing contract mismatches; Phase 5 catches integration-level protocol issues

Refactoring Decision Criteria (Wave 9)

• Module >500 lines: consider splitting
• 3+ modules share similar code: extract common module
• Naming inconsistency flagged by rtl-critic: rename pass
• Missing parameterization: add parameters for magic numbers
• Refactoring is selective — not all modules need it
• Refactor equivalence proof policy:
  • Cosmetic/style-only cleanup: lint + smoke simulation minimum
  • Any change touching combinational/sequential logic, reset, clock enable, or constraints intent: invoke equivalence-checker (RTL-vs-RTL) before Wave 10 gate

Wave 6a/6b Unit Test Scope

• Wave 6a (Tier 1 Smoke): Basic I/O stimulus, FSM state coverage, self-checking assertions. Quick pass/fail per module. Produces: sim/{module}/tb_{module}.sv
• Wave 6b (Tier 2 Unit — MANDATORY): Reference model comparison (DPI-C or file-based), uarch feature-level coverage with REQ-U-* tracing, per-feature result JSON. Deeper per-module verification. Extends: sim/{module}/tb_{module}.sv (adds reference comparison logic to existing Wave 6a TBs, does NOT replace them). Gate: Tier 2 gate per "Phase 4 Gate Criteria" below (single normative statement of thresholds and required {module}_unit_results.json fields).
• rtl-p4s-unit-test (standalone skill): Can still be invoked independently for ad-hoc Tier 2 testing outside the pipeline

AC Coverage at Wave 6b (Advisory)

When iron-requirements has structured acceptance_criteria (with ac_id) for a REQ-U-*: ac_ids should be populated for each unit test feature covering that requirement. This is advisory at P4 (not hard-block). Incomplete AC coverage is flagged as WARNING and deferred to P5 for closure. When no structured AC: existing req_ids gate applies unchanged.

Rapid-Impl to Full-Impl Transition

• rtl-p4-rapid-impl produces: lint-clean modules, module-level CDC pass, smoke functional pass, block sanity pass. State in p4-state.json.
• rtl-p4-implement adds: code review (Wave 4-5), Tier 1 smoke tests (Wave 6a), Tier 2 unit tests (Wave 6b), protocol checks (Wave 8), refactoring (Wave 9), integration gate (Wave 10), Stream B artifacts.
• Transition: Full-impl orchestrator should detect existing lint-clean modules from rapid-impl and skip Waves 1-3 for those modules. Detection via: rtl/*/*.sv exists + lint passes.

Phase 4 Sub-Skills Integration

• rtl-p4s-bugfix: Used in Wave 5 for review-driven fixes, and Wave 6 for test-driven fixes
• rtl-p4s-refactor: Used in Wave 9 for code quality improvements
• rtl-p4s-unit-test: Tier 2 testing (used in Wave 6b for mandatory per-module verification; also available standalone)
• rtl-lint-check: Used in Waves 2-3 and after any code modification

Escalation & Stop Conditions

• Module still has lint errors after 3 fix rounds → escalate to rtl-architect for design review
• Module fails code review after 3 review→fix iterations → escalate to rtl-architect for structural redesign
• uarch spec is ambiguous for a module → pause that module, flag to user, continue others
• Unit test fails after 3 debug→fix→re-sim iterations → escalate to waveform-analyzer + rtl-architect
• CDC FAIL after 2 fix rounds → escalate to cdc-reviewer for synchronization strategy
• CDC FAIL where root cause is clock source/clock gating/clock mux relationship ambiguity → escalate to clock-architect (in addition to cdc-reviewer)
• Protocol FAIL after 2 fix rounds → escalate to protocol-reviewer for interface redesign
• Functional coverage review FAIL with >3 missing REQs → pause, flag to user (potential uarch spec gap)

Phase 4 Gate Criteria

ALL of the following must be true before Phase 5:

RTL Files:

• rtl//.sv exists for every block in docs/phase-3-uarch/
• rtl/filelist_{module}.f exists for every module
• rtl/filelist_top.f exists and includes all module filelists

Lint:

• All files pass verilator --lint-only -Wall with zero errors
• No module blocked after 3 lint fix rounds

Synthesizability (HARD — Wave 3.5):

• reviews/phase-4-rtl/{module}-synthesizability.md exists for every module with verdict PASS
• Wave 10 Stream-B synth smoke (stream-b-synth-estimate.md) reports zero CRITICAL (inferred latches / unmappable constructs) — gate-blocking, not advisory

Code Review:

• All modules reviewed by rtl-critic (Wave 4)
• All modules REVIEW_PASS (0 critical/major findings)
• Per-module review reports at .rat/scratch/phase-4/
• No module blocked after 3 review→fix iterations

Test Plan:

• sim/{module}/{module}_test_plan.md exists for every module (Wave 0 Step 0b)

Unit Test (Tier 1 + Tier 2):

• sim/{module}/tb_{module}.sv exists for every module
• All Tier 1 smoke tests PASS (sim/{module}/{module}_results.txt)
• All Tier 2 unit tests PASS (sim/{module}/{module}_unit_results.json with ref_mismatches=0)
• Tier 2 coverage meets thresholds: FSM >= 50%, line >= 60%
• Every feature entry in unit_results.json has req_ids populated (REQ-U-* tracing)
• Functional coverage: func_coverage.covergroups_defined >= 1 per module
• Codec conformance PASS (if H.264/H.265 decoder — Step 5a)

AC Coverage (advisory):

• ac_ids populated in unit_results.json when acceptance_criteria exist (advisory)

CDC:

• All multi-domain modules CDC-checked (Wave 7)
• All multi-domain modules CDC_PASS
• Single-domain modules explicitly marked as skip

Protocol:

• All bus-interface modules protocol-checked (Wave 8)
• All bus-interface modules PROTOCOL_PASS
• No-interface modules explicitly marked as skip

Refactoring:

• Flagged modules refactored (Wave 9)
• Equivalence verified for all refactored modules
• For logic/clock/reset-impact refactors: equivalence-checker report exists (RTL-vs-RTL)

Integration + Gate:

• Basic integration smoke test PASS
• rtl-critic functional coverage verdict is PASS
• Every REQ-NNN implemented in at least one RTL module
• reviews/phase-4-rtl/functional-completeness.md saved
• reviews/phase-4-rtl/design-review.md saved
• reviews/phase-4-rtl/lint-report.md saved

Stream B (content quality verified):

• docs/phase-4-rtl/stream-b-sva-skeletons.md saved — contains property/assert per module
• docs/phase-4-rtl/stream-b-cdc-preliminary.md saved — references clock domain names from clock-domain-map.md
• docs/phase-4-rtl/stream-b-tb-skeletons.md saved — references REQ- tag per module + contains test function/task
• docs/phase-4-rtl/stream-b-synth-estimate.md saved (synthesis smoke test — no inferred latches, no unmappable constructs)

Naming Conventions:

• All port names use i_/o_/io_ prefix (NOT suffix _i/_o)
• All clocks: clk or {domain}_clk — NOT clk_i, clk_sys
• All resets: rst_n or {domain}_rst_n — NOT rst_ni
• All instances: u_ prefix, generates: gen_ prefix
• logic only — no reg/wire keywords
• always_ff/always_comb — no bare always
• Parameters: UPPER_SNAKE_CASE, localparam: L_ prefix, types: snake_case_t

Summary:

• docs/phase-4-rtl/phase-4-summary.md generated