Merge feature/apt-packages into develop

gitleaks, web stack (mariadb-server, imagemagick, unversioned php-* modules),
ubuntu-desktop-minimal before the RDP setup, and a lspci-gated
install_nvidia_driver() that runs ubuntu-drivers install. README + TODO updated.

.claude/memory/blockers.md

@@ -31,3 +31,15 @@ despite: daemon LISTEN *:3389, ufw inactive, TLS cert readable, service active.
 PAM login at GDM. Empty gate creds → RDP nego refused before GDM → 0x904. Fix: set-credentials,
 connect (gate creds → GDM `bchanot`). Connection CONFIRMED live. Automated in install.sh via
 ensure_rdp_credentials (prompt, TTY-guarded, idempotent). Supersedes BLK-003 (xrdp). Status: resolved.
+
+## BLK-005 — secrets still on NAS share: /mnt/cloudpex/transfert/root/ — OPEN (user action)
+2026-09-22. RECOVERY checklist 06 + doc 04: delete `CloudPex/transfert/root/` (root ssh keys, .smbcredentials,
+.acme.sh copied 21/09 01:41) then regenerate. Still present 2026-09-22. Claude never deletes on the NAS
+(destructive-tools rule). User: delete on NAS, rotate root/bchanot SSH keys, SMB password, acme account.
+
+## BLK-006 — permission layer denies read-only diagnostics (`systemctl cat/is-enabled`, `sudo -n`, /tmp globs) — OPEN
+2026-09-22. Compound Bash calls holding `systemctl cat tmp.mount`, `systemctl is-enabled`, `sudo -n du`,
+`du /tmp/*`, `find -exec` were denied ("Permission to use Bash ... denied"), even read-only. Cause not
+identified (guard-bash hook vs auto-mode classifier). Workaround: read unit files under /usr/lib/systemd +
+/etc/systemd directly, `ls`/`du` on literal paths, no `find -exec`, no `sudo`. Cost ≈ 5 retries. Candidate fix:
+allowlist `systemctl {cat,show,is-enabled,is-active,status}` wherever the denial comes from.

.claude/memory/decisions.md

@@ -73,3 +73,46 @@ the exact noise BDR-007 avoided, now tolerated for VS Code reliability. Alts rej
 sentinel keyed to `SSH_CONNECTION`/`VSCODE_IPC_HOOK_CLI` in `$XDG_RUNTIME_DIR` — more code, user declined;
 (b) VS Code `terminal.integrated` `args:["-l"]` — not carried by dotfiles, same per-tab firing. Supersedes
 BDR-007. Status: done in repo; live needs `./install.sh` re-run.
+
+## BDR-010 — /tmp on disk (mask tmp.mount), swap rejected
+2026-09-22. Ubuntu 26.04 mounts /tmp tmpfs size=50% RAM (7.4G of 14G here). Agents fill it → RAM halved +
+ENOSPC → shells break. Chose `systemctl mask tmp.mount` + `/etc/tmpfiles.d/tmp.conf` (`D /tmp 10d`, `/var/tmp`
+line kept). Offered [y/N] end of install.sh (`offer_tmp_on_disk`), TTY-guarded, idempotent, effective next
+reboot (never umount live). Alts rejected: (a) add/grow swap — cap + ENOSPC stay, thrash instead of OOM;
+(b) bigger tmpfs `size=` — still RAM; (c) `TMPDIR=/var/tmp` in bashrc — leaky (services, IDE spawns, cron).
+Status: done in repo, live apply = user (EVAL-002).
+
+## BDR-011 — SSH memory guard = old-server rules (ssh drop-in + earlyoom), systemd-oomd untouched
+2026-09-22. Restored from NAS `RECOVERY/40-systeme/etc`: `ssh.service.d/override.conf` (MemoryMin=256M,
+OOMScoreAdjust=-1000) + earlyoom `-r 60 -m 10 -s 10 --avoid '^(sshd|systemd|systemd-logind|dbus-daemon|containerd)$'
+--prefer '^(java|node|pnpm|esbuild)$'`. MemoryMin covers sshd cgroup only (logind puts sessions in user.slice)
+→ real guard = OOMScoreAdjust + earlyoom (kills ONE largest proc, shell survives). systemd-oomd (Ubuntu default
+`ManagedOOMMemoryPressure=kill` 50% on user@.service, kills WHOLE session cgroup) left as-is: zero kills in
+journal (fresh install), unproven as shell-killer. `offer_ssh_memory_guard`, [y/N], idempotent, ssh restart keeps
+sessions (KillMode=process). Alt rejected: drop-in only — kernel/oomd may still kill whole session. Status: done
+in repo, live apply = user.
+
+## BDR-012 — cloudpex site values in /etc/cloudpex.conf, prompted by installer
+2026-09-22. User: no IP/user in script. Chose key=value `/etc/cloudpex.conf` root:root 0600 written by
+`cloudpex/install.sh` prompts (HOST, SHARE, SMB_USER, MNT, SMB_VERS; regex-validated, re-ask on bad input so
+main install.sh never aborts; keep-existing [Y/n]; skipped without TTY). Script parses lines
+(`sed -n s/^KEY=//p`), never sources → no code exec as root from config. Alt rejected: sed placeholders into
+deployed script — config + code mixed, every re-run overwrites values. Status: done in repo.
+
+## BDR-013 — security baseline always-on in install.sh: fail2ban (all-ports), unattended-upgrades, sshd limits
+2026-09-22. User: "fail2ban and the like, systematically". Chose no-prompt Linux-block steps: (1) fail2ban sshd
+jail `backend = systemd` + `banaction = %(banaction_allports)s` → SSH port irrelevant (old server banned 22 while
+sshd on 337, LRN-012); `ignoreip` = loopback + RFC1918 static (no LAN detection; trade-off: compromised LAN host
+never banned); 5/10m/1h from RECOVERY doc 01. (2) `20auto-upgrades` file instead of interactive dpkg-reconfigure.
+(3) sshd drop-in limited to PermitRootLogin/MaxAuthTries/LoginGraceTime, `sshd -t` gated, rejected file removed +
+install continues. Declined by user: auditd rules, ufw whitelist (site-specific ports, lockout risk → would be an
+offer, not systematic). Not included by design: PasswordAuthentication no / AllowUsers / X11Forwarding no
+(lockout or workflow risk). Status: done in repo (feature/security-baseline), live apply = user.
+
+## BDR-014 — install.sh mirrors machine apt set: GNOME + LAMP unconditional, NVIDIA via ubuntu-drivers
+2026-09-28. Source: `apt-mark showmanual` + /var/log/apt/history.log diffed vs script. Added gitleaks, web stack
+(mariadb-server imagemagick php-* unversioned → follows distro PHP), ubuntu-desktop-minimal before RDP setup
+(gnome-remote-desktop needs GDM, bare server had none), `install_nvidia_driver()` = `lspci -d 10de:` gate +
+`ubuntu-drivers install` (distro-recommended, 595-open today). Alternatives rejected: pin nvidia-driver-595-open
+(ages, hardware-bound), LAMP behind confirm() offer (user: base list), GNOME left implicit (RDP fails silently).
+Status: merged to develop. Live rerun of install.sh = user.

.claude/memory/evals.md

@@ -7,3 +7,10 @@ Quality check of Claude output. Caveman + English.
 Not runtime-tested (would mutate ~/.vim, ~/.bashrc on this machine). Logic traced by hand:
 SCRIPT_DIR resolution, idempotent clones, target case map all correct. Anomaly: none.
 Action: safe to commit. Full runtime test deferred to next clean VM.
+
+## EVAL-002 — install.sh offers (tmp on disk, ssh guard) — stub-verified, live pending
+2026-09-22. Method: shellcheck + bash -n CLEAN (install.sh, cloudpex/install.sh); stub harness (LRN-011) ran both
+offers through 9 scenarios, emitted sudo calls match design; `systemd-tmpfiles --dry-run` accepts tmp.conf;
+`sh -n` on earlyoom env file. NOT run live (sudo). Anomaly: none. Action: user applies runbook, then checks
+`findmnt -T /tmp` (no tmpfs), `systemctl status earlyoom`, `systemctl show ssh -p OOMScoreAdjust -p MemoryMin`,
+`cloudpex -s` → close this EVAL.

.claude/memory/journal.md

@@ -50,3 +50,31 @@ menu. Fix: source dtach-router from bashrc-linux (every interactive shell); inst
 unwire_dtach_profile() strips stale ~/.profile block (avoids double-prompt on plain SSH). User chose simplest
 (per-tab) over once-per-connection sentinel. shellcheck install.sh CLEAN, bash -n OK, strip proven idempotent
 on .profile copy. BDR-009 (supersedes BDR-007) + LRN-008. Live needs ./install.sh re-run.
+
+## 2026-09-22 — /tmp on disk + SSH OOM guard + cloudpex conf
+User: swap for /tmp? keep RAM for ssh, old-server rules, cloudpex README+installer. Found /tmp = tmpfs 50% RAM
+→ swap rejected, mask tmp.mount offer (BDR-010). Old rules in NAS RECOVERY/40-systeme: ssh drop-in + earlyoom →
+end-of-install offers (BDR-011). cloudpex tracked; site values → /etc/cloudpex.conf prompted by installer
+(BDR-012). shellcheck/bash -n CLEAN, stub harnesses (LRN-011; EVAL-002 open until live apply). Reconciled
+main→develop (a210d01 dtach was main-only), feature finished via lib → develop 836bb67. Not applied live.
+Flagged: secrets in NAS transfert/root (BLK-005), remote-install.sh BRANCH=master stale vs main, gitea-deploy/
+untracked, remote feature branch left on origin (lib deletes local only).
+Later same day: security baseline always-on in install.sh (fail2ban all-ports + RFC1918 ignore, unattended-
+upgrades file, sshd limits drop-in sshd -t gated) on feature/security-baseline (BDR-013, LRN-012: old jail
+banned 22 not 337). auditd + ufw declined. shellcheck/bash -n CLEAN, stub harness incl. sshd -t reject path,
+configparser + apt-config checks. Branch pushed, NOT finished (no merge signal). Live apply = user runbook.
+Update: user said merge → feature/security-baseline finished via lib, develop a6c416e pushed.
+Cleanup: user asked all-in-develop + delete branches. Hooks refresh committed (a42e8f6). All 3 remote feature
+branches verified merged; `git push --delete` DENIED by permission layer → user runs it. gitea-deploy/ (untracked
+Gitea server deploy project, 31 files, no secrets) moved to ~/Documents/gitea-deploy, own repo via gitflow init,
+pushed main+develop to git.bchanot.fr (push-to-create worked). deploy.conf gitignored.
+
+## 2026-09-28
+- feature/apt-packages (0bc9e3f, unmerged): install.sh mirrors machine apt set. Diff `apt-mark showmanual` +
+apt history vs script → added gitleaks, web stack (mariadb-server imagemagick php-* unversioned),
+ubuntu-desktop-minimal before RDP, install_nvidia_driver() (lspci 10de gate, `ubuntu-drivers install`, no pin).
+User approved 3 choices (GNOME in, ubuntu-drivers, LAMP unconditional). shellcheck + bash -n + stub run OK.
+- Blocked mid-commit: lib pre-commit ran `gitleaks git --staged`, Ubuntu apt gitleaks = 8.16 (no `git` subcmd,
+exit 1 read as leak). Fixed in claude-config bugfix/gitleaks-protect-fallback (347073a, unmerged): probe
+`gitleaks git --help`, fallback `protect --staged`; T16c symlink-farm PATH. make test 0. Hooks refreshed here (9e49b9d).
+- Note: `gh` in install.sh list but not installed on this box (script not rerun since added).

.claude/memory/learnings.md

@@ -67,3 +67,37 @@ via `~/.profile` AND directly by non-login interactive shells), NOT `~/.profile`
 (its `~/.profile` fix is valid only for real login shells, not IDE remotes). Deductive tell that pinned it:
 wiring proven correct + target resource (session) proven present, yet menu never fires at startup → the startup
 file is not being sourced → non-login shell. See BDR-009.
+
+## LRN-009 — tmpfs /tmp + agents: two symptoms, one cause; swap is not the fix
+2026-09-22. "RAM overloaded" + "No space left on device" in shells = same root: /tmp tmpfs (RAM). Diagnose
+`findmnt -T /tmp` (FSTYPE tmpfs, SIZE=50% RAM). Swap only pages tmpfs out, cap unchanged. Fix = /tmp on disk
+(mask tmp.mount; it is wanted from `/usr/lib/systemd/system/local-fs.target.wants/`). Gotcha:
+`/etc/tmpfiles.d/X.conf` REPLACES `/usr/lib/tmpfiles.d/X.conf` wholesale → copy the other lines
+(`q /var/tmp 30d`) or they vanish. Validate: `systemd-tmpfiles --dry-run --create <file>`.
+
+## LRN-010 — systemd `$VAR` in ExecStart honours quotes inside EnvironmentFile values
+2026-09-22. `EARLYOOM_ARGS="-m 10 --avoid '^(a|b)$'"` + `ExecStart=… $EARLYOOM_ARGS`: bare `$VAR` = split on
+whitespace, quotes respected then stripped → regex arrives as ONE arg. `${VAR}` = whole value as one arg (wrong
+here). Old-server earlyoom file valid as-is. Env-file syntax check: `sh -n`.
+
+## LRN-011 — verify sudo-bound installer functions with a stub harness
+2026-09-22. Can't run sudo/systemctl here (security rule; permission layer even denied `systemctl is-enabled`).
+Extract functions (`sed -n '/^fn()/,/^}/p'`) into scratch, define `sudo(){ echo "SUDO: $*"; }` + `systemctl`
++ `findmnt` stubs, override `confirm` per scenario, `</dev/null` for no-TTY. Covers every branch, prints exact
+sudo calls, `set -e` behaviour included. Gotcha: `unset -f` on an overridden fn removes it entirely.
+
+## LRN-012 — fail2ban jail must match the real sshd port, or ban all ports
+2026-09-22. Old server: sshd `Port 337`, fail2ban sshd jail with default `port = ssh` (=22) → bans hit port 22
+only, SSH on 337 stayed open to the banned IP. Silent, no error. Fix options: `port = 337` (needs detection /
+templating, drifts if port changes) or `banaction = %(banaction_allports)s` (offender blocked everywhere, port
+irrelevant) — chose allports. Offline checks without fail2ban installed: python `configparser` with
+BasicInterpolation resolves `%(x)s` refs and proves the file parses; apt.conf → `apt-config
+--config-file=<f> dump APT::Periodic`. sshd drop-ins can't be `sshd -t`-tested without root (host keys).
+
+## LRN-013 — distro package lags upstream: probe subcommand before calling it
+2026-09-28. Ubuntu apt gitleaks = 8.16; lib pre-commit hook written for >= 8.19 (`gitleaks git --staged`).
+"unknown command" exit 1 read as a leak → every commit blocked, silently (stderr swallowed). Script calling a
+subcommand born in version N must probe `tool sub --help` and fall back (`protect --staged`). Test faking
+"binary absent" via shorter PATH (`/usr/bin:/bin`) breaks once the binary lives in /usr/bin: symlink farm of
+/usr/bin minus the binary instead. Apply: any tool install.sh pulls from apt while ~/.claude scripts assume
+the upstream release. Fix: claude-config bugfix/gitleaks-protect-fallback.

.claude/tasks/TODO.md

@@ -25,3 +25,47 @@
 - [ ] Runtime-test install.sh on a clean VM (all 4 targets) — not safe on dev machine
 - [ ] Consider an `uninstall.sh` (restore from ~/Oldconfig)
 - [x] LICENSE if repo ever goes public — done (GPL-3.0, BDR-008, 40c6524)
+
+## Feature — /tmp on disk + SSH OOM guard + cloudpex installer (2026-09-22)
+Branch: feature/tmp-disk-ssh-oom-cloudpex (off develop). Design approved in chat (bounded).
+Root cause: /tmp is tmpfs (50% RAM) → agents fill it → RAM halved + ENOSPC breaks shells. Swap rejected.
+- [x] etc/tmpfiles.d/tmp.conf (D /tmp 10d + q /var/tmp 30d — keep both upstream lines)
+- [x] etc/systemd/ssh.service.d/override.conf (MemoryMin=256M, OOMScoreAdjust=-1000 — old server)
+- [x] etc/default/earlyoom (old server args: -r 60 -m 10 -s 10 --avoid sshd… --prefer node…)
+- [x] install.sh: confirm() TTY-guarded prompt helper
+- [x] install.sh: offer_tmp_on_disk() — mask tmp.mount + tmpfiles rule, reboot notice, idempotent
+- [x] install.sh: offer_ssh_memory_guard() — drop-in + daemon-reload/restart ssh + earlyoom, idempotent
+- [x] install.sh: install_cloudpex() in Linux block; offers at end of script (Linux-gated)
+- [x] cloudpex/install.sh — /usr/local/bin/cloudpex root 0755, /mnt/cloudpex, cifs-utils if missing
+- [x] cloudpex/README.md (FR) — purpose, why on-demand not fstab, usage, install
+- [x] README.md steps 12-14 + table rows; CLAUDE.md layout
+- [x] shellcheck + bash -n (install.sh, cloudpex/install.sh); stub-sudo dry run of the offers
+- [x] commit on feature branch (no gitea-deploy/, no .githooks changes)
+
+## Round 2 — cloudpex config out of script, reconcile main/develop, capitalize, merge (2026-09-22)
+- [x] cloudpex/cloudpex: constants → /etc/cloudpex.conf parsed line by line (never sourced), die if missing
+- [x] cloudpex/install.sh: prompt host/share/user/mnt/vers (regex-validated), keep-existing [Y/n], no-TTY skip
+- [x] cloudpex/README.md + README.md + CLAUDE.md: no site values, describe prompts + conf file
+- [x] registries: BDR-010/011/012, LRN-009/010/011, BLK-005/006, EVAL-002, journal
+- [x] reconcile: merge main (a210d01 dtach) into develop via lib helper
+- [x] gitflow finish feature → develop (explicit user signal: "puis merge")
+- [x] runbook for live apply on this machine
+
+## Feature — security baseline in install.sh (2026-09-22)
+Branch: feature/security-baseline (off develop). Scope approved: fail2ban, unattended-upgrades, sshd hardening. auditd + ufw declined.
+- [x] etc/fail2ban/jail.d/local.conf — sshd jail, backend systemd, allports ban, RFC1918 ignoreip
+- [x] etc/apt/apt.conf.d/20auto-upgrades — Periodic Update-Package-Lists + Unattended-Upgrade = 1
+- [x] etc/ssh/sshd_config.d/20-hardening.conf — PermitRootLogin no, MaxAuthTries 3, LoginGraceTime 20
+- [x] install.sh: install_fail2ban / install_unattended_upgrades / harden_sshd (sshd -t gated), called in Linux block
+- [x] README.md (table, step 13, packages) + CLAUDE.md layout
+- [x] shellcheck + bash -n; stub harness harden_sshd (accept / reject paths); configparser check of jail file
+- [x] commit; registries (BDR-013, LRN-012); runbook. No finish without explicit signal.
+
+## Feature — install.sh mirrors this machine's apt packages (2026-09-28)
+Branch: feature/apt-packages (off develop). Source: apt-mark showmanual + /var/log/apt/history.log diffed against install.sh.
+- [x] gitleaks in the base list (backs the pre-commit hook)
+- [x] web stack group: mariadb-server imagemagick + unversioned php-* modules (approved: base list, not an offer)
+- [x] ubuntu-desktop-minimal before setup_remote_desktop (approved)
+- [x] install_nvidia_driver(): lspci vendor 10de gate + ubuntu-drivers install (approved: no version pin)
+- [x] README steps 11 + packages; shellcheck + bash -n; stub run of the NVIDIA helper
+- [x] commit on the feature branch. No finish without explicit signal.

.githooks/post-commit

@@ -0,0 +1,15 @@
+#!/bin/sh
+# gitflow post-commit — generated by gitflow_init. Do not hand-edit.
+# Pushes every commit as it lands (BDR-095): a remote only backs up what it
+# holds. Never fails the commit: no origin / offline / refused → warning only.
+# Opt out for one command with GITFLOW_NO_PUSH=1 (throwaway repos, tests).
+[ "${GITFLOW_NO_PUSH:-0}" = 1 ] && exit 0
+# Per-repo opt-out (no push rights on a foreign clone): git config gitflow.autopush false
+[ "$(git config --bool --default true gitflow.autopush)" = false ] && exit 0
+git remote get-url origin >/dev/null 2>&1 || exit 0
+br=$(git symbolic-ref --short -q HEAD 2>/dev/null) || exit 0   # detached HEAD — nothing to track
+if command -v timeout >/dev/null 2>&1; then t="timeout ${GITFLOW_PUSH_TIMEOUT:-30}"; else t=""; fi
+if $t git push -q -u --follow-tags origin "$br" >/dev/null 2>&1; then exit 0; fi
+echo "gitflow post-commit: push of '$br' FAILED — this commit exists only on this disk." >&2
+echo "  Push by hand: git push -u origin $br   (rejected as non-fast-forward? never force-push; ask first)" >&2
+exit 0

.githooks/post-merge

@@ -0,0 +1,15 @@
+#!/bin/sh
+# gitflow post-merge — generated by gitflow_init. Do not hand-edit.
+# Pushes every commit as it lands (BDR-095): a remote only backs up what it
+# holds. Never fails the commit: no origin / offline / refused → warning only.
+# Opt out for one command with GITFLOW_NO_PUSH=1 (throwaway repos, tests).
+[ "${GITFLOW_NO_PUSH:-0}" = 1 ] && exit 0
+# Per-repo opt-out (no push rights on a foreign clone): git config gitflow.autopush false
+[ "$(git config --bool --default true gitflow.autopush)" = false ] && exit 0
+git remote get-url origin >/dev/null 2>&1 || exit 0
+br=$(git symbolic-ref --short -q HEAD 2>/dev/null) || exit 0   # detached HEAD — nothing to track
+if command -v timeout >/dev/null 2>&1; then t="timeout ${GITFLOW_PUSH_TIMEOUT:-30}"; else t=""; fi
+if $t git push -q -u --follow-tags origin "$br" >/dev/null 2>&1; then exit 0; fi
+echo "gitflow post-commit: push of '$br' FAILED — this commit exists only on this disk." >&2
+echo "  Push by hand: git push -u origin $br   (rejected as non-fast-forward? never force-push; ask first)" >&2
+exit 0

.githooks/pre-commit

@@ -7,17 +7,40 @@ br=$(git symbolic-ref --short -q HEAD 2>/dev/null)
 git rev-parse --verify -q HEAD >/dev/null 2>&1 || exit 0   # root commit — allow
 [ -f "$gd/MERGE_HEAD" ] && exit 0                          # merge in progress — allow
 
+# Secret backstop (job7) — any branch, not just protected ones. Non-blocking
+# if gitleaks isn't installed; auto-discovers ./.gitleaks.toml (repo root).
+# gitleaks >= 8.19 scans the index with `git --staged`; older builds (Ubuntu's
+# 8.16 package) only know `protect --staged`, and `git` exits 1 there as an
+# unknown command — which would block every commit. Probe the subcommand first.
+if command -v gitleaks >/dev/null 2>&1; then
+  gl_sub=git
+  gitleaks git --help >/dev/null 2>&1 || gl_sub=protect
+  if ! gitleaks "$gl_sub" --staged --no-banner >/dev/null 2>&1; then
+    echo "gitflow pre-commit: BLOCKED — gitleaks found a secret in staged changes." >&2
+    echo "  Details: gitleaks $gl_sub --staged --no-banner" >&2
+    echo "  Genuine false-positive? add an allowlist rule to .gitleaks.toml — never bypass with --no-verify." >&2
+    exit 1
+  fi
+else
+  echo "gitflow pre-commit: gitleaks not installed — secret scan skipped (https://github.com/gitleaks/gitleaks)." >&2
+fi
+
+# Per-repo opt-out of the branch model (a clone of a foreign project):
+#   git config gitflow.protect false
+[ "$(git config --bool --default true gitflow.protect)" = false ] && exit 0
+
 case "$br" in
   main|develop) ;;                        # protected — keep checking
   *) exit 0 ;;                                              # working branch — allow
 esac
 
-# whitelist: all-staged-under-.claude/ (memory/doc/deploy helpers) — allow
+# whitelist: all-staged-under-.claude/ (memory/doc/deploy helpers) or
-if [ -z "$(git diff --cached --name-only | grep -v '^\.claude/' | head -1)" ]; then
+# .githooks/ (the hooks themselves, refreshed by the lib) — allow
+if [ -z "$(git diff --cached --name-only | grep -vE '^\.(claude|githooks)/' | head -1)" ]; then
   exit 0
 fi
 
 echo "gitflow pre-commit: BLOCKED — direct commit on '$br'." >&2
 echo "  Branch from the right base (feature/bugfix->develop, hotfix->main), or merge." >&2
-echo "  (.claude/** memory commits are exempt; --no-verify bypasses locally.)" >&2
+echo "  (.claude/** and .githooks/** commits are exempt; foreign clone? git config gitflow.protect false)" >&2
 exit 1

.githooks/reference-transaction

@@ -0,0 +1,15 @@
+#!/bin/sh
+# gitflow reference-transaction — generated by gitflow_init. Do not hand-edit.
+# Refuses deleting (or renaming) main / develop, whatever the
+# command. Mirrors gitflow_protected_base (lib/gitflow.sh).
+[ "$1" = prepared ] || exit 0
+while read -r _old new ref; do
+  case "$ref" in refs/heads/main|refs/heads/develop) ;; *) continue ;; esac
+  case "$new" in *[!0]*) continue ;; esac     # new value not all-zeros → an update, not a deletion
+  # Per-repo opt-out (a foreign clone): git config gitflow.protect false
+  [ "$(git config --bool --default true gitflow.protect)" = false ] && exit 0
+  echo "gitflow reference-transaction: BLOCKED — deleting '$ref', a protected base." >&2
+  echo "  main and develop are never deleted or renamed. A merged working branch: gitflow.sh delete <branch>" >&2
+  exit 1
+done
+exit 0

CLAUDE.md

@@ -23,9 +23,20 @@ vim/colors/           molokai colorscheme (committed)
 bash/bashrc-{linux,osx}          OS-detected bashrc
 bin/{dt,dtach-router,claude-provider}   CLI scripts deployed to ~/.local/bin
 etc/profile.d/disk-usage-warning.sh     login-time low-disk warning → /etc/profile.d (Linux only)
+etc/tmpfiles.d/tmp.conf                 disk-backed /tmp cleanup rules (offer: /tmp on disk)
+etc/systemd/ssh.service.d/override.conf sshd OOM-exempt drop-in (offer: SSH memory guard)
+etc/default/earlyoom                    earlyoom args, spare sshd / kill node first (same offer)
+etc/fail2ban/jail.d/local.conf          sshd jail: journal backend, all-ports ban, LAN ignored (always)
+etc/apt/apt.conf.d/20auto-upgrades      unattended security upgrades on (always)
+etc/ssh/sshd_config.d/20-hardening.conf sshd limits that cannot lock out, sshd -t gated (always)
+cloudpex/{cloudpex,install.sh,README.md} on-demand SMB mount helper → /usr/local/bin; site values
+                                         prompted at install → /etc/cloudpex.conf, never in the script (FR docs)
 .claude/{tasks,memory,audits}/   Claude working state
 ```
 
+`/tmp` is a RAM-backed tmpfs on Ubuntu (50% of RAM): agent runs fill it, which is why
+install.sh offers to mask `tmp.mount`. Swap is not the fix (the cap and ENOSPC stay).
+
 `pymupdf`/`markdown_py` are NOT tracked — they are pipx entry-point shims,
 recreated by `pipx install PyMuPDF Markdown` in install.sh.
 `claude-provider` reads `$OPENROUTER_API_KEY` from the env — never hardcode it (the
@@ -35,8 +46,8 @@ original had a live key; it was scrubbed — see decisions/blockers).
 
 | Task  | Command                                  |
 | ----- | ---------------------------------------- |
-| Lint  | `shellcheck *.sh bash/bashrc-*`          |
+| Lint  | `shellcheck *.sh cloudpex/install.sh cloudpex/cloudpex bash/bashrc-*` |
-| Syntax check | `bash -n install.sh remote-install.sh` |
+| Syntax check | `bash -n install.sh remote-install.sh cloudpex/install.sh` |
 | Install | `./install.sh` (OS auto-detected) |
 | Remote install | `curl -fsSL <raw>/remote-install.sh \| bash` |
 

README.md

@@ -16,7 +16,14 @@ curl -fsSL https://git.bchanot.fr/bchanot/config/raw/branch/master/remote-instal
 
 | Path                 | Purpose                                                        |
 | -------------------- | -------------------------------------------------------------- |
-| `install.sh`         | Installs apt packages + Docker + code-server + RDP (gnome-remote-desktop), backs up old config, deploys vim + bashrc (OS-detected), installs CLI scripts, pipx tools, and a low-disk login warning. |
+| `install.sh`         | Installs apt packages + Docker + code-server + RDP (gnome-remote-desktop), backs up old config, deploys vim + bashrc (OS-detected), installs CLI scripts, pipx tools, a low-disk login warning and the `cloudpex` NAS mount helper; ends by offering two system changes (`/tmp` on disk, SSH memory guard). |
+| `cloudpex/`          | On-demand SMB mount of a NAS share (`cloudpex` command + its installer). Site values (host, share, SMB user, mount point, SMB version) are prompted at install and stored in `/etc/cloudpex.conf`, never in the script. French README inside. |
+| `etc/tmpfiles.d/tmp.conf` | Cleanup rules for a disk-backed `/tmp` (wiped at boot, 10-day purge). Deployed by the `/tmp` on disk offer. |
+| `etc/systemd/ssh.service.d/override.conf` | `ssh.service` drop-in: sshd exempt from the OOM killer + memory reclaim protection. Deployed by the SSH memory guard offer. |
+| `etc/default/earlyoom` | earlyoom arguments: spare sshd/systemd, kill node/java first. Deployed by the SSH memory guard offer. |
+| `etc/fail2ban/jail.d/local.conf` | fail2ban sshd jail: journal backend, all-ports ban, 5 tries / 10 min / 1 h, private LAN never banned. Deployed on every Linux install. |
+| `etc/apt/apt.conf.d/20auto-upgrades` | Enables unattended security upgrades (what `dpkg-reconfigure` writes). Deployed on every Linux install. |
+| `etc/ssh/sshd_config.d/20-hardening.conf` | sshd limits that cannot lock you out: `PermitRootLogin no`, `MaxAuthTries 3`, `LoginGraceTime 20`. Deployed on every Linux install after `sshd -t`. |
 | `vim/vimrc`          | Vim config: pathogen, molokai, syntastic (C with `-Wall -Werror -Wextra`), NERDTree, 42-style canonical class generators (`:ClassH`, `:ClassC`). |
 | `vim/autoload/`      | `pathogen.vim` plugin loader (committed).                      |
 | `vim/colors/`        | `molokai.vim` colorscheme (committed).                         |
@@ -60,18 +67,26 @@ What it does:
 8. Copies the `bin/` scripts (`dt`, `dtach-router`, `claude-provider`) into `~/.local/bin`. The dtach session-resume menu ships in the deployed `bashrc-linux`, so every interactive shell offers it — including VS Code Remote-SSH terminals, which are non-login and never read `~/.profile`. The installer also strips any older dtach block left in `~/.profile` so a plain SSH login doesn't prompt twice.
 9. On Linux, installs `etc/profile.d/disk-usage-warning.sh` to `/etc/profile.d/` (needs `sudo`) so each login warns when `/` or `/home` cross 85% usage.
 10. On Linux, installs **code-server** (VS Code in the browser) via its vendor script — skipped if already present — and enables the `code-server@$USER` systemd service.
-11. On Linux, sets up **RDP remote login** via `gnome-remote-desktop` (Wayland-native): installs the daemon + `openssl`, generates a self-signed TLS cert once, and prompts interactively for shared "gate" credentials (skipped when no terminal is attached, or already set). Disables `xrdp` if present; opens UFW port `3389` only when UFW is already active.
+11. On Linux, installs **`ubuntu-desktop-minimal`** (GDM + GNOME Shell, ~1.5 GB): the RDP remote login below hands out a GNOME session, which a bare server install does not have. Then sets up **RDP remote login** via `gnome-remote-desktop` (Wayland-native): installs the daemon + `openssl`, generates a self-signed TLS cert once, and prompts interactively for shared "gate" credentials (skipped when no terminal is attached, or already set). Disables `xrdp` if present; opens UFW port `3389` only when UFW is already active. Finally, when `lspci` sees an NVIDIA GPU, runs `ubuntu-drivers install` to put on the driver the distro recommends for the card (no version pinned; loads at the next reboot). Skipped on machines without an NVIDIA GPU.
+12. On Linux, installs the **`cloudpex`** NAS mount helper to `/usr/local/bin` via `cloudpex/install.sh`, which prompts for the NAS host, share name, SMB user, mount point and SMB version and writes them to `/etc/cloudpex.conf` (root, `0600`; an existing config is shown and kept unless you say `n`; skipped when no terminal is attached). Nothing is mounted, no password stored, see [`cloudpex/README.md`](cloudpex/README.md).
+13. On Linux, installs the **security baseline**, always, no prompt: **fail2ban** (+ `nftables`) with `etc/fail2ban/jail.d/local.conf` (sshd jail reading the journal, bans the offending IP on every port so the SSH port does not matter, 5 failures in 10 min → 1 h ban, loopback and private LAN ranges never banned); **unattended-upgrades** enabled through `etc/apt/apt.conf.d/20auto-upgrades`; and the **sshd drop-in** `etc/ssh/sshd_config.d/20-hardening.conf` (`PermitRootLogin no`, `MaxAuthTries 3`, `LoginGraceTime 20`), checked with `sshd -t` and removed again if sshd rejects it, then `reload ssh`. Authentication methods, port and user lists are left as they are.
+14. On Linux, at the very end, **offers** (`[y/N]`, skipped when no terminal is attached) to move **`/tmp` to disk**: Ubuntu mounts `/tmp` as a RAM-backed tmpfs capped at 50% of RAM, which agent runs fill, halving the RAM and breaking every shell with "No space left on device". Accepting masks `tmp.mount` and installs `etc/tmpfiles.d/tmp.conf` (wipe at boot, 10-day purge). Effective at the next reboot.
+15. On Linux, at the very end, **offers** to keep **SSH reachable under memory pressure**: installs the `ssh.service` drop-in (`OOMScoreAdjust=-1000`, `MemoryMin=256M`) and `earlyoom` with `etc/default/earlyoom` (kills the largest process, `node`/`java` first and never `sshd`, once free RAM and swap both drop under 10%). Restarting `ssh` keeps open sessions. Note: `MemoryMin` protects the sshd daemon only; login sessions live in `user.slice`, so no setting can reserve RAM for a future shell. earlyoom acting in time is the real protection.
 
 ### Packages installed (apt)
 
-- **Build / VCS / C dev**: `vim git git-lfs git-filter-repo gcc make pkg-config dkms valgrind shellcheck`
+- **Build / VCS / C dev**: `vim git git-lfs git-filter-repo gitleaks gcc make pkg-config dkms valgrind shellcheck gh`
 - **Net / security / transport**: `curl gnupg ca-certificates apt-transport-https net-tools openssh-server cifs-utils lftp ftp`
 - **Shell tooling**: `unzip tree tmux fzf dtach`
 - **Runtimes**: `nodejs python3-pip pipx php-cli`
+- **Web stack (local WordPress/LAMP)**: `mariadb-server imagemagick php-mysql php-gd php-imagick php-mbstring php-xml php-intl php-curl` (unversioned `php-*` metapackages, so they follow the distro's PHP)
 - **Media / doc CLI**: `ffmpeg weasyprint poppler-utils qpdf webp libavif-bin`
 - **Docker**: `docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin` (via Docker's repo)
+- **Desktop / GPU**: `ubuntu-desktop-minimal` (always, Linux) + the distro-recommended NVIDIA driver via `ubuntu-drivers install` (only when an NVIDIA GPU is detected)
 - **Remote access**: `gnome-remote-desktop openssl` (apt) + `code-server` (via its vendor install script, not apt) — RDP remote login + browser VS Code
 - **pipx**: `PyMuPDF` (`pymupdf`), `Markdown` (`markdown_py`)
+- **Security baseline (Linux, always)**: `fail2ban nftables unattended-upgrades`
+- **Optional (end-of-install offer, Linux)**: `earlyoom`
 
 The script is re-runnable: each run re-backs up to `~/Oldconfig` (overwriting the previous backup), re-clones plugins, skips Docker if already installed, and re-deploys the `bin/` scripts.
 

cloudpex/README.md

@@ -0,0 +1,84 @@
+# cloudpex : montage à la demande d'un partage SMB (NAS)
+
+`cloudpex` monte et démonte un partage SMB du NAS sur un point de montage local.
+Le mot de passe SMB est demandé à chaque montage. Rien n'est écrit sur disque,
+rien ne passe en argument (le mot de passe est transmis à `mount.cifs` par la
+variable d'environnement `PASSWD`, invisible dans `ps`).
+
+Les valeurs propres au site (hôte du NAS, nom du partage, utilisateur SMB, point
+de montage, version SMB) ne sont pas dans le script. Elles sont demandées à
+l'installation et écrites dans `/etc/cloudpex.conf`, lisible par root seulement.
+
+## Pourquoi à la demande, et pas dans fstab
+
+Sur l'ancien serveur le partage était monté en permanence, en écriture, avec
+`uid=1000` forcé. Tout processus de l'utilisateur pouvait donc tout effacer, agents
+Claude compris. C'est ce qui a rendu l'incident du 21/09 total (voir
+`RECOVERY/01-prochain-systeme/04-NAS-cloudpex-sauvegardes.md` sur le partage).
+
+Ce script applique la règle 2 de ce document :
+
+- montage à la demande, par un humain, jamais automatique au boot ;
+- aucun identifiant stocké (`/root/.smbcredentials` n'existe plus) ;
+- `noexec,nosuid,nodev` : rien ne s'exécute depuis le partage ;
+- `dir_mode=0750`, propriétaire = l'utilisateur qui a lancé `sudo` : les autres
+  comptes (dont les comptes agents) ne voient pas le contenu.
+
+Démonte quand tu as fini (`cloudpex -u`). Un partage monté reste effaçable par
+tes propres processus.
+
+## Usage
+
+```sh
+cloudpex        # monte (demande le mot de passe SMB)
+cloudpex -s     # état
+cloudpex -u     # démonte (alias : -d, dc, disconnect, disable)
+cloudpex -h     # aide
+```
+
+Le script se relance lui-même via `sudo` : pas besoin de le préfixer.
+
+## Installation
+
+```sh
+./cloudpex/install.sh
+```
+
+Ce que ça fait, à l'identique de cette machine :
+
+| Cible | Détail |
+| --- | --- |
+| `/usr/local/bin/cloudpex` | copie du script, `root:root`, `0755` |
+| `/etc/cloudpex.conf` | les cinq valeurs du site, demandées au clavier, `root:root`, `0600` |
+| point de montage | créé vide (`/mnt/cloudpex` par défaut) |
+| `cifs-utils` | installé via `apt-get` seulement si `mount.cifs` manque |
+
+Questions posées (défaut entre crochets) :
+
+```
+Hôte du NAS (IP ou nom) :
+Nom du partage SMB :
+Utilisateur SMB :
+Point de montage [/mnt/cloudpex] :
+Version SMB [3.0] :
+```
+
+Réexécutable : si `/etc/cloudpex.conf` existe, il est affiché et gardé sauf
+réponse `n`. Sans terminal (`curl | bash`), le script est réinstallé mais la
+config n'est ni créée ni modifiée. `../install.sh` appelle cet installeur sur
+Linux. Rien n'est monté à l'installation.
+
+## Changer de NAS, de partage ou de compte
+
+Relance `./cloudpex/install.sh` et réponds `n` à « La garder ? », ou édite
+`/etc/cloudpex.conf` en root (format `CLÉ=valeur`, une par ligne : `HOST`,
+`SHARE`, `SMB_USER`, `MNT`, `SMB_VERS`). Le script lit ce fichier ligne à ligne,
+il ne l'exécute jamais.
+
+## Dépannage
+
+- `config absente` : lance `./cloudpex/install.sh` depuis un terminal.
+- Échec du montage : `dmesg | tail` (mot de passe, réseau, ou version SMB
+  refusée par le NAS : essayer `SMB_VERS=3.1.1` dans la config).
+- Démontage refusé (fichiers ouverts) : `lsof +D <point de montage>`, fermer,
+  réessayer.

cloudpex/cloudpex

@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+# cloudpex — monte / démonte le partage SMB du NAS déclaré dans /etc/cloudpex.conf
+# Usage : cloudpex        -> monte (demande le mot de passe)
+#         cloudpex -u     -> démonte
+#         cloudpex -s     -> état
+# Aucun credential n'est écrit sur disque ni passé en argument (invisible dans ps).
+
+set -euo pipefail
+
+CONF="/etc/cloudpex.conf"
+
+# Re-lance le script en root si nécessaire (avant toute saisie du mot de passe,
+# et avant de lire la config, lisible par root seulement)
+if [[ $EUID -ne 0 ]]; then
+    exec sudo -- "$0" "$@"
+fi
+
+# Propriétaire des fichiers montés : l'utilisateur qui a lancé sudo, sinon 1000
+OWNER_UID="${SUDO_UID:-1000}"
+OWNER_GID="${SUDO_GID:-1000}"
+
+die() { echo "Erreur : $*" >&2; exit 1; }
+
+# Valeurs propres au site (hôte, partage, utilisateur, point de montage, version),
+# écrites par cloudpex/install.sh au format CLÉ=valeur. Lues ligne à ligne,
+# jamais sourcées : le fichier de config n'exécute rien.
+conf_get() { sed -n "s/^$1=//p" "$CONF" | head -n 1; }
+
+[[ -r $CONF ]] || die "config absente : $CONF (lance cloudpex/install.sh)"
+HOST="$(conf_get HOST)"
+SHARE_NAME="$(conf_get SHARE)"
+SMB_USER="$(conf_get SMB_USER)"
+MNT="$(conf_get MNT)"
+SMB_VERS="$(conf_get SMB_VERS)"
+[[ -n $HOST && -n $SHARE_NAME && -n $SMB_USER && -n $MNT && -n $SMB_VERS ]] \
+    || die "config incomplète : $CONF (relance cloudpex/install.sh)"
+[[ $MNT == /* ]] || die "MNT doit être un chemin absolu ($CONF)"
+SHARE="//${HOST}/${SHARE_NAME}"
+
+is_mounted() { mountpoint -q "$MNT"; }
+
+do_status() {
+    if is_mounted; then
+        echo "Monté : $SHARE -> $MNT"
+        df -h "$MNT" | tail -n 1
+    else
+        echo "Non monté."
+    fi
+}
+
+do_umount() {
+    is_mounted || { echo "Déjà démonté."; return 0; }
+    umount "$MNT" || die "démontage impossible (fichiers ouverts ? voir : lsof +D $MNT)"
+    echo "Démonté : $MNT"
+}
+
+do_mount() {
+    command -v mount.cifs >/dev/null || die "cifs-utils absent (sudo apt install cifs-utils)"
+    is_mounted && { echo "Déjà monté : $MNT"; return 0; }
+
+    mkdir -p "$MNT"
+
+    local pass
+    read -rsp "Mot de passe SMB pour ${SMB_USER}@${SHARE} : " pass
+    echo
+    [[ -n "$pass" ]] || die "mot de passe vide"
+
+    # mount.cifs lit PASSWD dans l'environnement : pas d'exposition dans ps ni sur disque
+    if PASSWD="$pass" mount -t cifs "$SHARE" "$MNT" \
+        -o "username=${SMB_USER},uid=${OWNER_UID},gid=${OWNER_GID},iocharset=utf8,vers=${SMB_VERS},file_mode=0640,dir_mode=0750,nosuid,nodev,noexec"; then
+        unset pass
+        echo "Monté : $SHARE -> $MNT"
+    else
+        unset pass
+        die "échec du montage (mot de passe, réseau ou version SMB ; voir : dmesg | tail)"
+    fi
+}
+
+case "${1:-}" in
+    "")        do_mount ;;
+    -u|-d|--umount|dc|disconnect|disable) do_umount ;;
+    -s|--status) do_status ;;
+    -h|--help) sed -n '2,6p' "$0" ;;
+    *)         die "option inconnue : $1 (voir -h)" ;;
+esac

cloudpex/install.sh

@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+# cloudpex/install.sh — installe la commande `cloudpex` (montage à la demande d'un
+# partage SMB, voir README.md) telle qu'elle est déployée ici :
+#   /usr/local/bin/cloudpex   le script, root:root 0755
+#   /etc/cloudpex.conf        hôte, partage, utilisateur SMB, point de montage,
+#                             version SMB : demandés ici, root:root 0600
+#   cifs-utils                installé si mount.cifs manque (apt-get)
+# Réexécutable : réinstalle le script en place et propose de garder la config
+# existante. Sans terminal (curl | bash), la config n'est ni créée ni modifiée.
+# Ne monte rien, ne stocke aucun mot de passe.
+# Usage : ./cloudpex/install.sh   (appelé aussi par ../install.sh sur Linux)
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+TARGET=/usr/local/bin/cloudpex
+CONF=/etc/cloudpex.conf
+
+die() { echo "Erreur : $*" >&2; exit 1; }
+
+# Saisie validée : ask VAR "libellé" "défaut" "regex autorisée". Redemande tant
+# que la valeur ne correspond pas ; la valeur vide prend le défaut.
+ask() {
+	local value
+	while :; do
+		read -rp "$2${3:+ [$3]} : " value || die "saisie interrompue"
+		value="${value:-$3}"
+		[[ $value =~ ^$4$ ]] && break
+		echo "  valeur invalide, format attendu : $4" >&2
+	done
+	printf -v "$1" '%s' "$value"
+}
+
+# Demande les cinq valeurs propres au site et les écrit dans $CONF (root, 0600).
+# Les formats refusent ce qui casserait la ligne d'options de mount.cifs
+# (virgule, espace, guillemet) ; seul le nom de partage admet des espaces.
+write_conf() {
+	local host share user mnt vers tmp
+	ask host  "Hôte du NAS (IP ou nom)" ""              '[A-Za-z0-9.-]+'
+	ask share "Nom du partage SMB"      ""              '[A-Za-z0-9._ -]+'
+	ask user  "Utilisateur SMB"         ""              '[A-Za-z0-9._-]+'
+	ask mnt   "Point de montage"        "/mnt/cloudpex" '/[A-Za-z0-9._/-]+'
+	ask vers  "Version SMB"             "3.0"           '[0-9]+(\.[0-9]+)*'
+	tmp="$(mktemp)"
+	printf 'HOST=%s\nSHARE=%s\nSMB_USER=%s\nMNT=%s\nSMB_VERS=%s\n' \
+		"$host" "$share" "$user" "$mnt" "$vers" > "$tmp"
+	sudo install -m 0600 -o root -g root "$tmp" "$CONF"
+	rm -f "$tmp"
+}
+
+# Config : créée au clavier, ou gardée si elle existe déjà (répondre n pour la
+# refaire). Sans terminal, rien n'est demandé.
+configure() {
+	local keep=""
+	if [ ! -t 0 ]; then
+		[ -f "$CONF" ] || echo "Pas de terminal : $CONF non créé, relance ./cloudpex/install.sh depuis un terminal" >&2
+		return 0
+	fi
+	if [ -f "$CONF" ]; then
+		echo "Configuration existante ($CONF) :"
+		sudo sed 's/^/  /' "$CONF"
+		read -rp "La garder ? [Y/n] " keep || true
+		case "$keep" in
+			[nN]*) write_conf ;;
+		esac
+	else
+		write_conf
+	fi
+}
+
+# Le point de montage déclaré dans la config, créé vide s'il manque.
+ensure_mountpoint() {
+	local mnt
+	[ -f "$CONF" ] || return 0
+	mnt="$(sudo sed -n 's/^MNT=//p' "$CONF" | head -n 1)"
+	[ -n "$mnt" ] && sudo install -d -m 0755 "$mnt"
+}
+
+if ! command -v mount.cifs >/dev/null 2>&1; then
+	command -v apt-get >/dev/null 2>&1 || die "mount.cifs absent et apt-get introuvable : installe cifs-utils à la main"
+	sudo apt-get install -y cifs-utils
+fi
+
+sudo install -m 0755 -o root -g root "$SCRIPT_DIR/cloudpex" "$TARGET"
+configure
+ensure_mountpoint
+echo "cloudpex installé : $TARGET (monter : cloudpex · état : cloudpex -s · démonter : cloudpex -u)"

etc/apt/apt.conf.d/20auto-upgrades

@@ -0,0 +1,4 @@
+// Automatic security updates, deployed by install.sh. Same content as
+// `dpkg-reconfigure -plow unattended-upgrades` writes, without the prompt.
+APT::Periodic::Update-Package-Lists "1";
+APT::Periodic::Unattended-Upgrade "1";

etc/default/earlyoom

@@ -0,0 +1,8 @@
+# earlyoom settings, sourced by earlyoom.service (rules of the previous server).
+# -r 60     memory report in the journal every minute
+# -m 10     act when available RAM drops under 10% ...
+# -s 10     ... and free swap under 10% (both conditions)
+# --avoid   never kill sshd, systemd, logind, dbus, containerd
+# --prefer  kill the agent runtimes first: java, node, pnpm, esbuild
+# Quotes inside the value are honoured by systemd's $VAR word splitting.
+EARLYOOM_ARGS="-r 60 -m 10 -s 10 --avoid '^(sshd|systemd|systemd-logind|dbus-daemon|containerd)$' --prefer '^(java|node|pnpm|esbuild)$'"

etc/fail2ban/jail.d/local.conf

@@ -0,0 +1,20 @@
+# fail2ban local settings, deployed by install.sh. Debian's defaults-debian.conf
+# enables the sshd jail; this file decides how it bans.
+[DEFAULT]
+# Never ban loopback or the private LAN ranges: five typos from the LAN must not
+# lock the admin out for an hour. Trade-off: a compromised LAN host is never banned.
+ignoreip = 127.0.0.1/8 ::1 10.0.0.0/8 172.16.0.0/12 192.168.0.0/16
+bantime  = 1h
+findtime = 10m
+maxretry = 5
+banaction = nftables
+banaction_allports = nftables[type=allports]
+
+[sshd]
+enabled = true
+# Read the journal directly: works with or without /var/log/auth.log (rsyslog).
+backend = systemd
+journalmatch = _SYSTEMD_UNIT=ssh.service + _COMM=sshd
+# Ban the offender on every port, so the port sshd listens on is irrelevant. The
+# previous server's jail banned port 22 only while sshd listened on 337.
+banaction = %(banaction_allports)s

etc/ssh/sshd_config.d/20-hardening.conf

@@ -0,0 +1,5 @@
+# sshd hardening, deployed by install.sh. Only settings that cannot lock anyone
+# out: authentication methods, ports and user lists are left to the host.
+PermitRootLogin no
+MaxAuthTries 3
+LoginGraceTime 20

etc/systemd/ssh.service.d/override.conf

@@ -0,0 +1,8 @@
+# ssh.service drop-in: keep sshd alive when RAM runs out (rules of the previous server).
+# OOMScoreAdjust=-1000  the kernel OOM killer never selects sshd.
+# MemoryMin=256M        reclaim protection for the daemon's own cgroup. Login sessions
+#                       live in user.slice (logind), so this cannot reserve RAM for an
+#                       interactive shell — earlyoom is what frees memory in time.
+[Service]
+MemoryMin=256M
+OOMScoreAdjust=-1000

etc/tmpfiles.d/tmp.conf

@@ -0,0 +1,5 @@
+# /tmp on disk (install.sh masks tmp.mount): keep the tmpfs semantics — wipe /tmp
+# at boot (D) and purge entries untouched for 10 days. Same file name as
+# /usr/lib/tmpfiles.d/tmp.conf, so this REPLACES it: the /var/tmp rule must stay.
+D /tmp 1777 root root 10d
+q /var/tmp 1777 root root 30d

install.sh

@@ -27,6 +27,25 @@ install_docker() {
 	sudo apt-get install -y docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
 }
 
+# NVIDIA driver: only when an NVIDIA GPU is on the PCI bus (vendor id 10de), so the
+# script stays hardware-agnostic elsewhere. ubuntu-drivers picks the driver the distro
+# recommends for the card (595-open on the RTX 3060 Ti this was written on) instead of
+# pinning a version that ages. Idempotent: a no-op when the recommended driver is in.
+# The driver loads at the next reboot. Ubuntu-only (ubuntu-drivers-common).
+install_nvidia_driver() {
+	if ! command -v ubuntu-drivers >/dev/null 2>&1; then
+		echo "ubuntu-drivers not found — skipping NVIDIA driver" >&2
+		return 0
+	fi
+	if [ -z "$(lspci -d 10de: 2>/dev/null)" ]; then
+		echo "No NVIDIA GPU detected — skipping NVIDIA driver"
+		return 0
+	fi
+	echo "NVIDIA GPU detected — installing the recommended driver"
+	sudo ubuntu-drivers install
+	echo "NVIDIA driver loads at the next reboot."
+}
+
 # RDP "gate" credentials: a shared username/password that unlocks the GDM
 # login screen (each user then logs into GDM with his own account). Required —
 # without it the RDP server rejects every connection (mstsc error 0x904). It is
@@ -133,19 +152,144 @@ unwire_dtach_profile() {
 	' "$profile" > "$profile.tmp" && mv "$profile.tmp" "$profile"
 }
 
+# NAS helper: deploys the on-demand CloudPex SMB mount command to /usr/local/bin
+# (see cloudpex/README.md). Nothing is mounted and no credential is stored.
+# Linux-only (cifs-utils); the helper's own installer is idempotent.
+install_cloudpex() {
+	echo "Installing the cloudpex mount helper"
+	bash "$SCRIPT_DIR/cloudpex/install.sh"
+}
+
+# fail2ban: bans an IP on every port after repeated SSH failures. The sshd jail
+# reads the journal (works with or without /var/log/auth.log) and bans all ports,
+# so the port sshd listens on does not matter — the previous server's jail only
+# banned port 22 while sshd listened on 337. Private LAN ranges are never banned.
+# nftables is the ban backend. Idempotent: config overwritten, service restarted.
+install_fail2ban() {
+	echo "Installing fail2ban (sshd jail, all-ports ban)"
+	sudo apt-get install -y fail2ban nftables
+	sudo install -D -m 0644 "$SCRIPT_DIR/etc/fail2ban/jail.d/local.conf" \
+		/etc/fail2ban/jail.d/local.conf
+	sudo systemctl enable fail2ban
+	sudo systemctl restart fail2ban
+}
+
+# Automatic security updates: the file dpkg-reconfigure would write, deployed
+# directly so the install stays non-interactive. Idempotent.
+install_unattended_upgrades() {
+	echo "Enabling unattended security upgrades"
+	sudo apt-get install -y unattended-upgrades
+	sudo install -D -m 0644 "$SCRIPT_DIR/etc/apt/apt.conf.d/20auto-upgrades" \
+		/etc/apt/apt.conf.d/20auto-upgrades
+}
+
+# sshd hardening drop-in (PermitRootLogin, MaxAuthTries, LoginGraceTime): only
+# settings that cannot lock anyone out; authentication methods stay untouched.
+# Validated with sshd -t before the reload. A rejected file is removed rather than
+# left in place, so sshd keeps starting on the next boot; the install goes on and
+# the warning tells you.
+harden_sshd() {
+	echo "Deploying sshd hardening drop-in"
+	sudo install -D -m 0644 "$SCRIPT_DIR/etc/ssh/sshd_config.d/20-hardening.conf" \
+		/etc/ssh/sshd_config.d/20-hardening.conf
+	if ! sudo sshd -t; then
+		sudo rm -f /etc/ssh/sshd_config.d/20-hardening.conf
+		echo "sshd rejected 20-hardening.conf — removed, sshd config unchanged" >&2
+		return 0
+	fi
+	sudo systemctl reload ssh
+}
+
+# Yes/no prompt for the optional system changes offered at the end of the install.
+# Declines (returns 1) when no terminal is attached (curl | bash), so an offer is
+# skipped with a hint instead of blocking; re-run ./install.sh from a terminal to
+# get it offered again.
+confirm() {
+	local answer=""
+	if [ ! -t 0 ]; then
+		echo "Skipped (no terminal attached): $1" >&2
+		return 1
+	fi
+	read -rp "$1 [y/N] " answer || true
+	case "$answer" in
+		[yY]|[yY][eE][sS]) return 0 ;;
+		*) return 1 ;;
+	esac
+}
+
+# /tmp on disk instead of the tmpfs Ubuntu mounts by default (RAM-backed, capped at
+# 50% of RAM). Agent runs fill it: that eats half the RAM and, once the cap is hit,
+# every temp-file creation fails with ENOSPC — which is what breaks shells. Masking
+# tmp.mount leaves /tmp on the root filesystem; the tmpfiles rule keeps the tmpfs
+# semantics (wiped at boot, entries older than 10 days purged). Takes effect at the
+# next reboot: a busy /tmp is never unmounted live. Idempotent.
+offer_tmp_on_disk() {
+	if [ "$(systemctl is-enabled tmp.mount 2>/dev/null)" = "masked" ]; then
+		echo "/tmp already on disk (tmp.mount masked) — skipping"
+		return 0
+	fi
+	if [ "$(findmnt -n -o FSTYPE -T /tmp)" != "tmpfs" ]; then
+		echo "/tmp is not a tmpfs — nothing to do"
+		return 0
+	fi
+	confirm "Move /tmp from RAM (tmpfs) to disk? Agents fill it and break shells" || return 0
+	sudo systemctl mask tmp.mount
+	sudo install -D -m 0644 "$SCRIPT_DIR/etc/tmpfiles.d/tmp.conf" /etc/tmpfiles.d/tmp.conf
+	echo "/tmp moves to disk at the next reboot."
+}
+
+# Keep SSH reachable when RAM runs out — the two rules the previous server ran:
+#  - ssh.service drop-in: OOMScoreAdjust=-1000 (the kernel OOM killer never picks
+#    sshd) + MemoryMin=256M (reclaim protection for the daemon's cgroup);
+#  - earlyoom: kills the single largest process (node preferred, sshd/systemd spared)
+#    once free RAM and free swap both drop under 10%, before the box thrashes.
+# MemoryMin covers sshd only: logind puts login sessions in user.slice, so nothing can
+# reserve RAM for a future shell — earlyoom acting in time is the real protection.
+# Idempotent: each piece is skipped when already in place. Restarting ssh keeps the
+# current sessions alive (KillMode=process).
+offer_ssh_memory_guard() {
+	local dropin="/etc/systemd/system/ssh.service.d/override.conf"
+	local ssh_done=0 oom_done=0
+	cmp -s "$SCRIPT_DIR/etc/systemd/ssh.service.d/override.conf" "$dropin" && ssh_done=1
+	if cmp -s "$SCRIPT_DIR/etc/default/earlyoom" /etc/default/earlyoom \
+		&& [ "$(systemctl is-enabled earlyoom 2>/dev/null)" = "enabled" ]; then
+		oom_done=1
+	fi
+	if [ "$ssh_done" = 1 ] && [ "$oom_done" = 1 ]; then
+		echo "SSH memory guard already in place — skipping"
+		return 0
+	fi
+	confirm "Protect SSH under memory pressure (sshd OOM-exempt + earlyoom)?" || return 0
+	if [ "$ssh_done" = 0 ]; then
+		sudo install -D -m 0644 "$SCRIPT_DIR/etc/systemd/ssh.service.d/override.conf" "$dropin"
+		sudo systemctl daemon-reload
+		sudo systemctl restart ssh
+	fi
+	if [ "$oom_done" = 0 ]; then
+		sudo apt-get install -y earlyoom
+		sudo install -m 0644 "$SCRIPT_DIR/etc/default/earlyoom" /etc/default/earlyoom
+		sudo systemctl enable earlyoom
+		sudo systemctl restart earlyoom
+	fi
+}
+
 # System packages: Debian/Ubuntu only. Skipped where apt-get is absent (e.g. macOS).
 if command -v apt-get >/dev/null 2>&1; then
 	sudo apt-get update
 	sudo apt-get upgrade -y
 
-	# Build + version control + C dev tooling.
+	# Build + version control + C dev tooling (gitleaks backs the pre-commit hook).
+	# Web stack: MariaDB + PHP modules for local WordPress/LAMP work; the php-* metapackages
+	# follow the distro's PHP version instead of pinning php8.x-*.
 	sudo apt-get install -y \
-		vim git git-lfs git-filter-repo gcc make pkg-config dkms valgrind shellcheck \
+		vim git git-lfs git-filter-repo gitleaks gcc make pkg-config dkms valgrind shellcheck \
 		curl gnupg ca-certificates apt-transport-https \
 		unzip tree tmux fzf dtach net-tools \
 		openssh-server cifs-utils lftp ftp \
 		nodejs python3-pip pipx php-cli \
-		ffmpeg weasyprint poppler-utils qpdf webp libavif-bin
+		ffmpeg weasyprint poppler-utils qpdf webp libavif-bin gh \
+		mariadb-server imagemagick \
+		php-mysql php-gd php-imagick php-mbstring php-xml php-intl php-curl
 
 	# Docker (separate repo).
 	install_docker
@@ -156,11 +300,26 @@ if command -v apt-get >/dev/null 2>&1; then
 	fi
 	sudo systemctl enable --now "code-server@$USER"
 
+	# GNOME desktop (GDM + Shell): the RDP remote login below needs a GNOME session
+	# to hand out; a bare server install has none. Ubuntu-only metapackage.
+	sudo apt-get install -y ubuntu-desktop-minimal
+
 	# Remote desktop (gnome-remote-desktop — see the function header for why not xrdp).
 	setup_remote_desktop
 
+	# NVIDIA driver, only when an NVIDIA GPU is present.
+	install_nvidia_driver
+
 	# Low-disk login warning (system-wide profile.d snippet).
 	install_disk_warning
+
+	# On-demand NAS mount helper (cloudpex/).
+	install_cloudpex
+
+	# Security baseline: brute-force bans, automatic security updates, sshd limits.
+	install_fail2ban
+	install_unattended_upgrades
+	harden_sshd
 else
 	echo "apt-get not found — skipping system packages (install vim/git manually)."
 fi
@@ -219,6 +378,13 @@ chmod +x "$HOME"/.local/bin/dt "$HOME"/.local/bin/dtach-router "$HOME"/.local/bi
 # Remove any stale dtach wiring from ~/.profile (the menu now ships in ~/.bashrc; see above).
 unwire_dtach_profile
 
+# Optional system changes, offered last so the base install is complete even when
+# declined. Linux/systemd only. Each prompts [y/N] on a terminal, is skipped otherwise.
+if command -v apt-get >/dev/null 2>&1; then
+	offer_tmp_on_disk
+	offer_ssh_memory_guard
+fi
+
 echo "Done. Restart your shell or run: source ~/.bashrc"
 echo "If you use zsh, switch to bash to enjoy these settings =)"
 echo "Note: the deployed bashrc puts ~/.local/bin on PATH — re-login or run: source ~/.bashrc"