[HugoXK]

The problem

The RRSI wiring (examples/rrsi/rrsi_policies.py, merged #254) re-derived two things the pinned upstream fixtures already provide: component attribution used a first-path-segment heuristic (skills/x/SKILL.md -> "skills") instead of the pinned components vocabulary (K / K_STR / normalize evidence rules -> "skill"), and the per-edit ledger was an in-memory list instead of the pinned history.History (durable JSONL L_t, merged #253). The fidelity claim in docs/algo-rrsi.md said the port is pinned; the wiring was not.

What this PR changes

• Component attribution through the pinned vocabulary. RRsiState.components_of now classifies each op via components.normalize / classify_diff: skills/x/SKILL.md is a skill edit, ToolRegistry(...) is client_tool, a pure prose edit is prompt. Novelty and pruning now count the components RRSI actually counts.
• The per-edit ledger is history.History (durable) when history_path= is given, with an in-memory twin for scratch runs and tests, pinned to the same summaries by a parity test so the twin cannot silently drift.
• RRsiProposalPolicy reads History.render / prune_set (with the actionable accepted edits to remove) / stall_flag.
• RRsiAcceptance scores novelty over History.incumbent_component_counts and records every decision via History.append_candidate.
• Checkpoint separation. The aggregator checkpoint now carries only the search state (round, S*, incumbent cost, progress); the per-edit evidence is the History's own file, so a resumed run reads the same evidence the loop conditioned on.
• stalled() uses the pinned history.stall_flag, which requires a full window of progress points (t >= w) — a stricter, correcter reading than the old tail-slice heuristic; the integration tests were updated to that semantics.

Fidelity / deviations

Behavior	Before	Now
skills/x/SKILL.md	"skills" (path segment)	"skill" (upstream K_STR)
ToolRegistry(...)	path segment	"client_tool"
pure prose edit	path segment	"prompt"
per-edit ledger	in-memory list in the checkpoint	history.History JSONL (durable)
stall	tail-slice heuristic	pinned stall_flag (t >= w)
resume	ledger lost	History's own file re-read on resume

Verification

Item	Result
Upstream vocabulary classifies skills/tool/memory/prompt correctly	2 tests
Memory and file History backends agree (yield_g/prune_set/render/counts)	1 test
Durable History survives reopen and resume	1 test
Checkpoint carries search state; History stays durable	1 test
A genuinely novel structural component carries a within-band candidate	1 test
Existing integration/history/selection/schedule suite	107 passed
Changelog fragment valid	exit 0

Files

5 files changed, +435, -159

Modified:

• examples/rrsi/rrsi_policies.py — upstream_components, _InMemoryHistory, History-backed RRsiState/RRsiProposalPolicy/RRsiAcceptance, build_rrsi_bundle(history_path=)
• tests/test_rrsi_integration.py — updated to the upstream vocabulary and History API (intent preserved)
• docs/algo-rrsi.md — the wiring now consumes the pinned fixtures

New:

• tests/test_rrsi_fidelity.py (+~150) — 6 fidelity tests
• changelog.d/331-rrsi-fidelity.md — fragment

[Birfy]

Verdict: APPROVE

This is the right fidelity closure: the wiring stops re-deriving what components / history already pin, and the docs/tests match the claim.

Correctness

• upstream_components → normalize(None, key\nvalue) is the right call site. Spot-checked skills/… → skill, ToolRegistry → client_tool, prose → prompt; novelty/K_STR now see the vocabulary Algorithm 2 actually counts.
• judge(..., history.incumbent_component_counts()) is a real fix vs the old attempts Counter. _novelty / components.novelty want accepted structural counts on the incumbent; feeding proposal attempts suppressed novelty after a rejected try.
• Checkpoint / History split is sound: aggregator payload keeps search state (round / S* / cost / progress); the JSONL is the durable ledger. Resume with the same history_path re-reads the evidence the loop conditioned on.
• stalled() → stall_flag(progress, len-1, w, δ) matches the pinned t >= w guard; integration tests correctly moved to a full window.

Design fit

• Reusing History / components instead of a parallel ledger is the port convention the module docstring promised.
• _InMemoryHistory as a scratch twin + parity test is a good compromise for tests that do not want a path.
• build_rrsi_bundle(history_path=...) is the right seam; docs/algo-rrsi.md updated accordingly.

Tests / verification (local)

Fresh clone at f79e936. 107 RRSI tests green (test_rrsi_fidelity 6, integration 22, history 31, selection 29, schedule 19). CI: build green; test matrix 3.9–3.13 still pending at review time.

Docs / changelog

Fidelity table and wiring section in docs/algo-rrsi.md match the code. Fragment is present and readable.

Non-blocking follow-ups

1. Changelog slug 331-rrsi-fidelity.md — open issue #331 is the unrelated Calibrator min_labels/min_per_stratum validation bug. Prefer renaming the fragment to track this PR (e.g. 333-…) or an issue you open for the fidelity closure, so the next release notes are not confused with #331.
2. Parity is incomplete on rounded fields. File History.append_candidate rounds delta_S/delta_C/S/C; _InMemoryHistory does not (and skips _finite). Clean values in test_memory_and_file_history_agree pass, but e.g. delta_S=0.123456789 / C=100.44 makes render diverge. Mirror the rounding (and ideally _finite) in the twin, or assert only the summary methods that stay equal.
3. Nits: unused import json in tests/test_rrsi_fidelity.py; unused field import in rrsi_policies.py; mid-module from .components import K as _K_ALL could sit with the other imports.
4. Outcome spelling: wiring writes lowercase "accepted"/"rejected" while MEASURED_OUTCOMES documents ACCEPTED/REJECTED/LOST. Harmless today (append does not validate), but aligning would reduce future drift.

No merge from this review (routine policy).

[Birfy]

Two additional P2 findings from independent source review at f79e9361c13db355f11b1d99a95577eb2f91da8b, beyond the existing follow-ups. Both concern the opt-in RRSI example bundle. The scenarios below are source-derived and proposed regressions, not locally executed reproductions (project execution remains on hold).

1. Default checkpoint/restore loses the history used for decisions. checkpoint / restore now omit all history even when history_path=None selects the in-memory backend. A fresh default bundle restores S*/round/progress but loses accepted component counts, pruning evidence and proposal history. The existing end-to-end resume test uses this default and only checks S*. Proposed regression: record an accepted skill, round-trip the default state, and compare the next flat-score/flat-cost skill candidate with uninterrupted execution (w_n=1). The resumed state should not award novelty again. Please preserve in-memory history in the checkpoint, or automatically arrange durable history for checkpointed runs; the explicit same-file path case does not cover this default.

2. Raw replacement text bypasses the prose-only classifier. upstream_components passes key + '\n' + value to normalize, while text_only recognizes changed lines only with +/- markers. A SingleSlot prompt replacement such as Do not mention ToolRegistry in the final answer therefore matches client_tool, although no tool machinery was added. That false structural tag supplies novelty to a flat candidate. Proposed regression: create this prompt via SingleSlot.to_diff, require component prompt, and require no novelty-based acceptance at equal score/cost. Please supply actual edit/context semantics rather than treating structural words in replacement prose as structural edits.