From c8c72c24aaddd06064bc1e4db490a481892e5f1f Mon Sep 17 00:00:00 2001
From: Bastien Chanot <git@bchanot.fr>
Date: Mon, 6 Jul 2026 01:06:54 +0200
Subject: [PATCH] =?UTF-8?q?chore(memory):=20capitalize=20=E2=80=94=20LRN-0?=
 =?UTF-8?q?04,=20EVAL-001,=20BDR-006=20update=20note,=20journal=202026-07-?=
 =?UTF-8?q?06?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .claude/memory/decisions.md |  1 +
 .claude/memory/evals.md     |  9 +++++++++
 .claude/memory/journal.md   |  6 ++++++
 .claude/memory/learnings.md | 10 ++++++++++
 4 files changed, 26 insertions(+)

diff --git a/.claude/memory/decisions.md b/.claude/memory/decisions.md
index a1ce7cb..3896892 100644
--- a/.claude/memory/decisions.md
+++ b/.claude/memory/decisions.md
@@ -108,6 +108,7 @@ rules:
 - **Why**: SEC-1 tour finding — stock `nginx:*-alpine` runs its master as root inside the container. Unprivileged image + port 8080 removes the root master; the rest shrinks blast radius. BDR-004's "port 80 / nginx:1.27-alpine / HSTS omitted at container" no longer matched the tree.
 - **Supersedes**: BDR-004 — topology unchanged (native front proxy → container on loopback); only the base image, internal port, and uid change.
 - **Reference**: `Dockerfile`, `docker-compose.yml`, `nginx.conf`, `nginx-security-headers.conf`. Fix commit `ba13d69`; drift caught by tour REC-1 (`.claude/audits/TOUR.md`, run 2026-07-05-2).
+- **Update 2026-07-06**: base bumped `1.28-alpine` → `1.30-alpine` digest-pinned (nginx/1.30.3) — 1.28 branch retired, CVE-2026-42945 fixed 1.30.1+ only, no backport. Decision unchanged (unprivileged base, 8080, uid 101). Commit `1aa97f0`, tour 2026-07-05-3 REC-2.
 
 ---
 
diff --git a/.claude/memory/evals.md b/.claude/memory/evals.md
index 80575df..c7a6665 100644
--- a/.claude/memory/evals.md
+++ b/.claude/memory/evals.md
@@ -21,6 +21,15 @@ rules:
 
 | ID | Date | Output | Action |
 |----|------|--------|--------|
+| EVAL-001 | 2026-07-06 | /tour run 2026-07-05-3 (3 it., converged) + residual closure | keep |
+
+## EVAL-001 — /tour run 2026-07-05-3 + residual closure pass
+
+- **Date**: 2026-07-06
+- **Output**: 3-iteration tour (security/clean/reconcile/doc, converged at bound) + closure of all 10 residuals on owner GO. Commits `1aa97f0`/`613bfc0`/`2f5e51a` + follow-up.
+- **Method**: oracle-based — semgrep ×3 (deterministic PASS), PDF render-hash (LRN-003) for behavior-preserving proofs, docker oracles (build, nginx -t, header/dotfile/gzip/healthcheck curls), headless-browser screenshots 375+1440 (index font trim), brace counts, CSP-hash pinned==computed.
+- **Anomalies**: (1) cso add-on caught a HIGH (base-image CVE) two same-day semgrep-only tours missed — gstack was OFF then → LRN-004. (2) Fresh clean sweeps surfaced new info-tier nits each iteration (N1–N4 at it2) — convergence needed explicit reporting threshold in it3 prompt; bound of 3 did its job. (3) Session limit killed both it3 agents mid-flight — SendMessage transcript-resume recovered both, zero re-audit gap.
+- **Action**: keep
 
 <!-- Append entries below. Template:
 
diff --git a/.claude/memory/journal.md b/.claude/memory/journal.md
index 5ce64ad..3216086 100644
--- a/.claude/memory/journal.md
+++ b/.claude/memory/journal.md
@@ -41,3 +41,9 @@ rules:
 - Closed all 5 residuals on owner GO: CLN-6 aria-hidden CTA arrows (`607124a`), CLN-7/8 palette conformance (5 off-palette colors → tokens, PDF regen render-verified, `ede7576`), CLN-9 geo aligned landing→CV = Nantes relocation (`f515875`).
 - Decided: BDR-006 (hardened container, supersedes BDR-004 infra), BDR-007 (geo canonical = Nantes, supersedes BDR-003 geo).
 - Branch chore/tour-2026-07-05-2 finished → develop + pushed.
+
+## 2026-07-06
+
+- Tour 2026-07-05-3 finished (session-limit pause mid-it3, agents transcript-resumed): CONVERGED 3 it. SEC-1 HIGH fixed — base 1.28→1.30-alpine, CVE-2026-42945 (`1aa97f0`); clean F1-F8 (`613bfc0`, −54 lines, render-hash proven); README synced (`2f5e51a`).
+- All 10 residuals closed on owner GO: Google Fonts trimmed both files (CV render-hash identical, index browser-verified 375+1440), CV date chips French + en-dash, CV tags → 999px pills, contact-grid no-ops dropped, nginx dotfile block reordered first, PDF gzip dropped, compose healthcheck deduped (image healthcheck verified healthy), CLAUDE.md white-family + CV-typography exception documented, BDR-006 annotated (1.30 bump).
+- LRN-004 (pinned base = frozen CVE exposure) + EVAL-001 (tour verdict) logged. Branch chore/tour-2026-07-05-3 → develop on owner GO (this session).
diff --git a/.claude/memory/learnings.md b/.claude/memory/learnings.md
index 3fd4fc1..ec32425 100644
--- a/.claude/memory/learnings.md
+++ b/.claude/memory/learnings.md
@@ -22,6 +22,7 @@ rules:
 | LRN-001 | 2026-05-15 | certbot --nginx matches `server_name`, not filename | nginx + certbot on multi-site VPS |
 | LRN-002 | 2026-05-17 | PIL supersample ×8 + Lanczos = clean icon antialiasing | Python stdlib icon generation |
 | LRN-003 | 2026-07-05 | Prove CSS cleanup behavior-preserving via before/after PDF render-hash | weasyprint / paged-media PDF projects |
+| LRN-004 | 2026-07-06 | Digest-pinned base image = frozen CVE exposure; SAST can't see it | any Dockerfile with pinned FROM |
 
 ---
 
@@ -49,3 +50,12 @@ rules:
 - **Pattern**: To confirm a CSS/HTML edit is truly behavior-preserving on a project whose deliverable is a weasyprint PDF: render a baseline PDF from the pre-edit HTML, apply the edit, regenerate, then compare (a) `pdftotext | sha256` and (b) per-page `pdftoppm -r 150 -png | sha256`. Text-hash alone misses `font-size`/color changes — the render-hash catches them. Identical render-hash = provably no visual change; and since weasyprint output is deterministic, an unchanged render yields a byte-identical PDF → nothing new to commit.
 - **Context**: tour clean phase on `bchanot-cv` removed dead CSS (`.reveal.d6`, `position:running()`, no-op `box-shadow`, dead `.skills-grid font-size`). Render-hash matched on both pages → proven before commit `30b0e44`. The same tooling later confirmed the intentional palette edit DID change the render (expected), distinguishing dead-code removal from real visual change.
 - **Future application**: Any weasyprint / paged-media project where you must tell "dead code removal" (must render identically) apart from "intended visual change". General trick: verify a refactor by hashing the rendered artifact, not the source.
+
+---
+
+## LRN-004 — Digest-pinned base image = frozen CVE exposure; SAST can't see it
+
+- **Date**: 2026-07-06
+- **Pattern**: Digest pin freezes image bytes → also freezes vulnerabilities. Pin correct at audit time can be HIGH same day: upstream retires stable branch, security batch lands only on newer branches, no backport. semgrep/SAST floor scans code, blind to base-image CVE freshness. Complementary posture pass required: base branch EOL status (endoflife.date) + vendor security advisories, every audit.
+- **Context**: bchanot-cv tour 2026-07-05-3. `nginx-unprivileged:1.28-alpine` digest-pinned as SEC fix in morning run; same evening cso posture add-on flagged HIGH — 1.28 branch retired, CVE-2026-42945 (rewrite-module overflow) fixed 1.30.1+/1.31.1+ only. Two intervening semgrep-only tours saw nothing (gstack OFF → no cso). Bump commit `1aa97f0`.
+- **Future application**: Any Dockerfile `FROM x@sha256:…` → security audit must include EOL + advisory check on the pinned branch, not just SAST. gstack ON → cso add-on covers it; OFF → manual endoflife.date + vendor advisory check.