git clone https://github.com/fluffos/nt7 → nt7

Source: fluffos/nt7 submodule at 04a1a75e (leftover 689: .lpc + §9 formatted on that repo; mudlib root is the repo root; filename-extension slices widened for .lpc). In-game name 王者归来 / NTLIB V7. Slug nt7, number 163, port 40211.

leftover 689: work/ is the fluffos/nt7 submodule

hosting is fluffos-upstream. work/ is now the fluffos/nt7 gitlink (pin 04a1a75e8f137138a1da21aae7c204572d872339), not a second formatted fork. config.fluffos mudlib directory still points at work/ (this repo's root *is* the mudlib). patches/ stays empty — the rename, §9 format, .lpc slices, and static→nosave keywords live on that fluffos-org repo. log_file("static/...") path literals were not rewritten (convert_lib's \bstatic\b also hits those; they stay "static/"). The catalog-only dead #ifndef __SENSIBLE_MODIFIERS__ shim in include/globals.h was not landed upstream. help/ntbook encoding sidecar stayed upstream-as-shipped. Three pre-existing lexerrors were left untouched: quest/skybook/update.lpc, inherit/condition/damo.lpc, d/dragon/outlet.lpc. PRs #3/#4/#5 were already on 53dce3a1 before this rename. Catalog-only overlay: data/login|user for fluffos / Mud@2026wiz (regular) + Mud@2026 (admin/recovery), including fluffos.package.o. Native 巫师休息室 (/d/wizard/ wizard_room) look/score/quit on fluffos / Mud@2026wiz (浮云, (boss)). score said 还没有出生呐 — by design (wizard skips the birth ceremony). Driver debug.log and work/log/ log_error were this boot (2026-09-12 08:17); unused-local compile warnings only. error_handler was not written (no caught runtime error). Zip pack: 29918 .lpc, overlay seed, no .git/ store. Full choose/washto/古村 registration was already verified at onboard.

earlier snapshot note (superseded)

Onboard convert was from 53dce3a1. Leftover 689 landed the catalog LPC delta on fluffos/nt7 and pinned that SHA.

• Source: git clone https://github.com/fluffos/nt7 (王者归来, "泥潭7"). Adapted per AGENTS.md §2.3 (already-git-hosted, already-FluffOS- adapted source), not a raw archive extraction — mudlib root IS the repo root (config.ini: mudlib directory : .). Repo history goes back to a nt7 for fluffos v2019 initial commit plus a handful of community fixes (调整登录，增加GBK端口, 修正代码, 修正部分BUG from contributor oiuv/steyeu) and, most recently, two commits from this project's own author fixing the exact Bad argument 4 to EFUN message() bug (§6.4-ish/message() varargs-exclude class) and seeding the fluffos admin account (§1.5) — i.e. this specific GitHub repo had already been independently worked on by this project's maintainer before being folded into the monorepo here. Both of those pre-existing fixes were verified still present/intact in the cloned source and left as-is.
• Encoding: verified, not assumed — every .c/.h file in the raw clone decodes as valid UTF-8 (0/30,337 failures via a full-tree Python UTF-8 decode check). scripts/convert_lib.sh run anyway per §2.3's "run it unconditionally" lesson (UTF-8 as the source encoding arg, so the whole pass is a no-op on content, just the mechanical .c→.lpc rename + literal-reference fixups): 29,920 .lpc files, 8,757 literal .c" refs fixed, 166 files' static→ nosave, 2 lossy-conversion files (help/ntbook, cosmetic help text). Repo-only cruft (.github/, .vscode/, scripts/pack_for_web.sh, a stray git_status dump file) stripped from work/ — not part of the mudlib itself.
• Port: 40211 (pre-assigned).

Status: native + WASM playable, full registration verified both ways

Lineage vs. the existing NT/nitan/Lonely family (AGENTS.md §2.1/§11)

Genuinely distinct branch, not a duplicate of any existing lib. Spot-checked core files (adm/kernel/master.lpc, adm/daemons/ logind.lpc, adm/daemons/securityd.lpc, adm/daemons/named.lpc, adm/daemons/chinesed.lpc) against nitan170911, nitan6, nt1, nt6, nt6nitan6win, xfbhh, and hhsj: every file differs substantially (hundreds to ~900 diff lines even under whitespace- insensitive diff -bw, out of files 150–1,850 lines long) — this is NOT the "byte-identical modulo reformatting" shape that predicts a wholesale fix-port (contrast: nt7's feature/dbase.lpc byte content vs. hhsj's post-reformat copy WAS identical modulo indentation style, confirming shared engine ancestry, but the higher-level daemon files have each evolved independently since the fork). Directory layout (adm/, clone/, cmds/, d/, data/, feature/, inherit/, kungfu/, quest/, u/, template/ …) and world content (d/ dongtian/, family/sect system, data/business bsave format) closely mirror hhsj's, and securityd.lpc's own file header credits "Doing Lu for ULTRA HELL" (the same "Doing"/hell lineage credit chain documented elsewhere in AGENTS.md §11) — strong evidence of a shared distant ancestor, not a recent fork of any single already-onboarded lib.

The single biggest payoff of checking lineage first: the family's signature architecture bug (AGENTS.md §7.15, "the nitan set/query/ dbase architecture bug") is ALREADY FIXED upstream here, and does not need porting. §7.15 describes the NT/nitan/Lonely lineage's classic trap: bare set()/query()/delete()/addn() simul_efun calls resolving this_object() to the SIMUL_EFUN OBJECT (wrong) rather than the caller, requiring a two-part fix (real local set/query in feature/dbase.lpc + a previous_object()-based simul_efun fallback). nt7's feature/dbase.lpc has NO local set/query/delete functions at all (same shape as the historically-BROKEN state on nitan170911/nitan6/hhsj/xfbhh/nt6/nt6nitan6win/nt1 before this project's own fixes) — but nt7's simul_efun shim (adm/kernel/simul_efun/ntefun.lpc, guarded #ifndef __PACKAGE_DBASE__, which is always true on this driver) correctly uses previous_object() (the CALLER) rather than this_object() (the simul_efun itself) to resolve the redirect target:

varargs mixed query(string prop, object ob) {
    if (!objectp(ob)) ob = previous_object();
    return ob->query_db(prop);
}

This is architecturally correct and was verified live (registration, washto, score all read/write player properties correctly). Since every bare call anywhere in the whole tree routes through this ONE correctly-implemented simul_efun (there being no local override anywhere to create the split-resolution shape §7.78 describes), the downstream §7.78 (mixin-file bare-call mis-resolution) and §7.79 (addn/addn_temp misdirection) bug classes don't apply here either — addn's shim uses the same previous_object() pattern. check_legal_name also already uses a real Unicode pcre_match(str, "^\\p{Han}+$") check (§8.1's "byte-range Chinese-detection" bug class does NOT apply — the GBK-byte-oriented is_chinese2() variant still exists in the same file but is unreachable dead code, confirmed via corpus-wide grep, left alone). Net effect: this branch appears to have either independently fixed, or never actually shipped, the exact bug class that was this project's single biggest recurring finding across the rest of this lineage — a useful confirmation that the bug is a fork-time regression in SOME branches, not baked into the shared upstream design itself as §7.15's own text once concluded.

Bugs found + fixed (programming/driver-compat only, per AGENTS.md §6)

1. §7.80-shaped bug, confirmed present, fixed: adm/daemons/ eventd.lpc's create() builds its event-file list with map_array(event_name, (: $1[0..<3] :)) to strip the .lpc extension — same off-by-one as the corpus-wide §7.80 sweep (str[0..<n] keeps len-n+1 chars on this driver, so stripping a 4-char suffix needs [0..<5], not [0..<3]). This lib wasn't part of that sweep (never onboarded before), so it still had the bug. Fixed [0..<3] → [0..<5]. Filed upstream, bundled with the 29 sibling sites found in round two below, as PR #3. 2. §4.3 string-literal collision, self-inflicted by the mechanical static→nosave conversion, fixed: 150 files/282 call sites had log_file("static/...", ...)-style path literals rewritten to "nosave/... by convert_lib.sh's blind \bstatic\b sed (the original source, confirmed via raw/, has zero "nosave/ strings anywhere — this is 100% conversion-tool fallout, not an upstream bug). Reverted every "nosave/ → "static/ across the whole tree. Not upstream-worthy (audited for the fluffos-org-PR queue, 2026-08-31): this is self-inflicted by this collection's own static→nosave conversion pass, not present in the upstream source at all — no PR filed. 3. #ifndef __SENSIBLE_MODIFIERS__ shim in include/globals.h, confirmed DEAD (not a live bug on this driver), removed for cleanliness anyway: the classic AGENTS.md §4.3 "#define protected static" shim was present (#define protected nosave), but this driver's LPC preprocessor already defines __SENSIBLE_MODIFIERS__ unconditionally (confirmed via the driver's own predefines dump), so the #ifndef branch was always compiled out — protected was already the real keyword, not aliased. Removed the dead shim block rather than leave a misleading comment; verified via a clean recompile (no new warnings/errors), and confirms feature/ command.lpc's protected nomask int command_hook(...) (§8.3a) was never actually broken here either — no fix needed for that one. Not upstream-worthy (audited for the fluffos-org-PR queue, 2026-08-31): a dead-code cleanup with no functional-bug signature, not a genuine bug fix — no PR filed. 4. New WASM-only bug, found + fixed, root-caused via bisection: the very FIRST connection under the WASM build always failed logon() silently (driver log: new_conn_handler: logon() on object clone/user/login#0 has failed, the user is disconnected., with ZERO LPC-level error/trace anywhere — safe_apply() just returns null) — reproducibly, at the exact same point, regardless of what was typed. Bisected with temporary write() markers spliced into logon(): everything up through chinese_number(max_usr) printed fine; the very next call, TIME_D->replace_ctime(time1), never even entered the target function body. Root cause: TIME_D (adm/daemons/timed.lpc) is listed in adm/etc/preload and DOES compile clean during boot preload (no warnings/errors at all), but is NOT actually resident (find_object(TIME_D) returns 0) by the time of the very first connection — and a bare call_other() (TIME_D->replace_ctime(...)) to a not-yet-loaded object, which transparently triggers an implicit load_object() and works fine under the NATIVE driver, silently fails the whole logon() call under the WASM build instead (confirmed empirically: an EXPLICIT catch(load_object(TIME_D)) immediately before the call fixes it completely, and every following line — including replace_ctime()'s own body — then executes normally). This looks like a genuine WASM-specific driver-level gap (implicit load-on-call_other not reliable there), not a mudlib logic bug, but per AGENTS.md §1.3's "mudlib-side guards are encouraged" policy, fixed at the mudlib level: if (!find_object(TIME_D)) catch(load_object(TIME_D)); immediately before the TIME_D->replace_ctime(time1) call in logind.lpc's logon(). Did NOT attempt to preemptively guard every other daemon call in this ~30k-file mega-lib against the same theoretical class — this was the one call site actually exercised and confirmed broken on the critical first-connection path; note for a future WASM pass on this lib (or a sibling) if another first-connection daemon call turns up the same symptom. Filed upstream, bundled with fix #5 below, as PR #5. 5. Unrelated hardening applied alongside #4: logind.lpc's logon() also had an UNCONDITIONAL DATABASE_D->db_count_user() call (real MySQL db_connect()) feeding a purely cosmetic "N registered players" banner line — every other DATABASE_D-> call in the same file is already correctly gated behind #ifdef DB_SAVE (disabled in this checkout's include/unixconf.h/winconf.h), so this one call was inconsistent with the file's own convention. Fixed to match: #ifdef DB_SAVE ... #else reg_usr = 0; #endif. Confirmed via bisection this was NOT the actual cause of bug #4 (removing it alone didn't fix the WASM hang) — kept anyway since it's a genuine, independent correctness/consistency fix (and doubly so under WASM, where PACKAGE_DB isn't compiled in at all, per the WASM build's own CMakePresets.json). Filed upstream bundled with fix #4, see PR #5.

Interactive verification

Native (~/src/fluffos/build-debug/src/driver config.fluffos, scripts/mudclient.py): full flow scripted end-to-end multiple times — id → confirm → surname/given-name (real Chinese, e.g. 秦风/秦峰/秦枫) → admin password → confirm → play password → confirm → gender → notice board → lands in /d/register/regroom ("泥潭注册室") → reg <email> completes the stub email-registration step (mail-sending is commented out upstream; registered flag just gets set) → moves to /d/register/entry ("生命之谷", meets 盘古/Pangu by name) → choose 1-4 → washto <str> <int> <con> <dex> (13–30 each, ≤80 total) auto- completes the "投胎" (reincarnation) step and moves the character straight into /d/city.../世界之树 in the Ancient Village (no separate born <city> command needed once washto succeeds — confirmed the game's own flow, not a bug: do_washto() calls the equivalent of do_born() internally). look/score both produce correct, fully-rendered output (character sheet shows the exact str/int/con/dex values from washto, correct gender/personality/age, zero mojibake). quit triggers the same "new account must stay online 30 minutes or it's deleted" policy already documented as intentional design on hhsj — confirmed by completing it with y: clean "账号...档案被 删除" message, save file actually removed, zero debug.log errors throughout any of the runs.

WASM (scripts/wasm_client.js against ~/src/fluffos/build-wasm/src): same full flow, same script, same result — registration through choose/washto/world-entry/score/quit all produce identical correct output to the native run, after fix #4 above. Before that fix, EVERY WASM connection attempt failed at the exact same point (silently, right after the "已经执行了N秒" uptime banner) regardless of what was typed — this was the actual §1.3/§1.4 WASM playability blocker for this lib, not an IP/socket/pcre-package issue like the family's other usual WASM traps.

Not fixed (deliberately, out of scope)

• Cosmetic compile warnings throughout the whole tree ("Unused local variable", "Number of arguments to X disagrees with previous definition", "Non-void functions must return a value", "Expression has no side effects") — pre-existing lazy-compile noise common to virtually every lib in this corpus, not touched per this project's scope discipline (programming bugs with an actual error signature, not lint-level noise).
• is_chinese2() in adm/kernel/simul_efun/chinese.lpc still has the old GBK-byte-range is_chinese() shape (§8.1's classic bug) — but it's genuinely dead code (zero callers anywhere in the tree, confirmed via grep), so left alone per this project's "don't fix unreachable code" convention.
• lpcc_check.sh's full-tree batch compile-sweep was started but KILLED partway through (RSS had reached ~15.7GB / >58% of host RAM on a shared multi-agent host after ~12 minutes, still climbing, no sign of finishing soon) — per AGENTS.md §10.4's explicit mega-lib guidance ("on mega-libs the boot + interactive test is the sufficient verification gate; the sweep is nice-to-have"), relied on the clean native boot (zero fatal errors, only cosmetic warnings) plus the full live registration/play verification above instead of the batch sweep.

Admin account seeding (AGENTS.md §1.5)

adm/etc/wizlist already shipped fluffos (boss) (seeded upstream by this project's own author in the pre-existing 254262ec commit, before this pass even started). No save file existed for it yet, so registered it for real through the normal flow (this lib's registration requires BOTH an admin/recovery password and a distinct regular login password — sending the same value for both is rejected): id fluffos, name 浮云 (surname 浮, given name 云 — picking the same character twice for both, e.g. "浮浮", trips a same-surname-name re-confirmation branch in get_name() that isn't a bug, just needs the joined full name typed back; used two different characters instead), admin password Mud@2026, regular login password Mud@2026wiz (matching this project's own established convention for this specific repo, from the 254262ec commit message). Since fluffos already resolves to (boss) via wizlist, enter_world()'s own startroom logic correctly routes it straight to WIZARD_ROOM (/d/wizard/wizard_room) instead of through the normal player birth-ceremony rooms — confirmed correct behavior, not a bug (wizard accounts are expected to skip reg/ choose/washto). Verified live: reconnect with the regular password succeeds, update /adm/daemons/timed reports "成功！". score reports "还没有出生呐" (not yet "born") for this account, since the normal birth-ceremony rooms that set the born property are never visited by a wizard-status login — did not chase this further (out of scope, likely by design for a staff-only account that isn't meant to use score).

Deep functional test (round two, 2026-08-27)

Full §10.7 playthrough (never previously done on this lib) against a freshly-rebuilt ~/src/fluffos/build-debug/src/driver, native only. Also checked all of AGENTS.md's standing cross-cutting bug-pattern list (§7.121, §8.3a, §7.112, §7.118, §7.122–§7.136, §7.139, §7.141– §7.148) via systematic grep — see per-pattern notes below. No sibling NT/nitan-lineage bug classes applied here beyond what's already documented above (§7.15 architecture bug already fixed upstream, per the earlier onboarding writeup).

Housekeeping note, not an nt7 bug: found an unrelated stray driver process (PID bound to port 40211, cwd in a completely different lib sanguozhi, reparented to init) squatting on this lib's assigned port before boot could even start — killed by exact PID before proceeding. Unrelated to nt7's own code; flagged here only so a future session knows this can happen on a shared host.

Bug found and fixed #1 — §7.80-shaped filename-slice bug, 29 more call sites in the kungfu/skill tree

explode(__FILE__, "/")[<1][0..<3] (and the dirs[<1][0..<3] variant) appears in 27 individual kungfu/skill/**/*.lpc files plus the two shared mixins inherit/meskill/skill_model_{weapon,unarmed}.lpc and the shared header kungfu/skill/force.h — the exact same str[0..<n] off-by-one this project already fixed in this very lib's adm/daemons/ eventd.lpc (.lpc is 4 characters, [0..<3] only strips 2, so "guiyuan-tunafa.lpc" yields "guiyuan-tunafa.l" instead of "guiyuan-tunafa"). This project's original onboarding pass only grepped for eventd.lpc's own exact line shape and never turned up this second, much larger family of call sites in the same lib. Two confirmed-live consequences:

• inherit/meskill/skill_model_weapon.lpc's practice_skill() calls me->query_skill(skname, 1) with the corrupted id, so this always looks up a nonexistent skill and returns 0 — silently breaking skill- level checks for the player-invented-martial-arts feature (cmds/ skill/invent.lpc, SKILL_MODEL_WEAPON/SKILL_MODEL_UNARMED templates).
• kungfu/skill/force.h's valid_public() (included by 9 top-tier force/internal-energy skill files, e.g. guiyuan-tunafa.lpc) builds a can_skill self-exclusion list containing the file's OWN corrupted id, so the self-exclusion check never actually matches the player's own already-partially-learned copy of that same skill — every learn attempt past the very first one hits "你不散掉归元吐呐法，如何能 修习归元吐呐法。" (an absurd "you must give up X before you can learn X" self-rejection), permanently capping every one of these 9 force skills at whatever their first learn happened to reach.

The remaining ~20 kungfu/skill call sites compute the same corrupted id but only pass it to SCBORN_D->valid_learn()/valid_perform() — a prerequisite-checking daemon whose file (adm/daemons/scbornd.lpc) does not exist anywhere in this archive at all (confirmed via corpus-wide grep for scbornd/SCBORN_D). Every such call unconditionally fails regardless of the id's correctness, so fixing the slice has no observable behavior change at those specific call sites — flagging this as a separate, out-of-scope missing-content gap (inventing what a whole prerequisite daemon should do would be a content/design decision, not a programming fix) rather than chasing it further. Fixed the slice arithmetic at all 29 sites anyway ([0..<3] → [0..<5]), since it's the same unambiguous mechanical bug independent of what currently consumes the result. Filed upstream, bundled with the eventd.lpc instance above, as PR #3.

One additional match, adm/daemons/skillsd.lpc's real valid_perform( object me, string file) (the actively-used SKILLS_D, as opposed to the unreferenced dev-sandbox copy at u/lonely/skillsd.lpc), had already had its own [0..<3] slice commented out and replaced with a bare dirs[<1] (no stripping at all) by a prior author — checked this against its real caller (inherit/skill/skill.lpc's perform_action(), which passes an already-extension-less path via perform_action_file()) and confirmed the no-slice form is actually CORRECT for that call site's real argument shape (the one other caller that would pass a .lpc-suffixed path is fully commented out, dead code). Left untouched.

Verified live: bai wu bo (拜师) then learn wu bo unarmed 1 correctly raised a fresh character's unarmed skill (skills showed 基本拳脚 (unarmed) - 不堪一击 6/14 after one learn call) — this exercises the general skill-learning path, not the specific 9 force-skill files, since reaching those requires much deeper progression; the force.h self-exclusion fix was verified by direct code-level tracing (confirmed the corrected id now matches what member_array() checks against) rather than a live multi-level force skill grind, given time budget — documented here as verified-by-code- reading rather than fully live-verified for that specific consequence.

Bug found and fixed #2 — a missing-content-factory null check silently truncates the crontab scheduler on every single boot

100%-reproducible on every fresh boot (not a random-roll fluke): adm/daemons/timed.lpc's init_crontab() reads adm/etc/crontab (140 lines) and does find_object(table[1]) || load_object(table[1]) on every referenced object to verify it resolves — one of the very first entries (line 120) references /u/redl/cangku (a wizard's personal treasure room, target of 6 华山论剑/牛人三部曲 tournament-scheduling crontab lines). Loading it for the first time runs its create(), which does EQUIPMENT_D->create_dynamic("", 60, 600)->move(this_ object()) in a loop — an unguarded §7.147-shaped chain. Worse, create_dynamic() ITSELF (adm/daemons/equipmentd.lpc:1368) has no null check either: ob = TEMPLATE_D->create_object(filename, obj_type, temp_status); can return 0 (whenever the randomly-rolled obj_type/ level combination doesn't resolve to a real shipped template file), and the very next statements unconditionally do set(..., ob) and ob->set_color(color) — a bare call_other() on int(0), which throws uncaught: *Bad argument 1 to EFUN call_other() Expected: object, string, array, Got: int(0) (debug.log, reproduced on every single boot in this session, both before and after the fix was verified absent).

Severity beyond the immediate crash: the error is only caught by an OUTER catch() several frames further up the call stack (logind. lpc's pre-existing catch(load_object(TIME_D)), from this lib's own bug #4 above) — which means the entire unwind aborts init_crontab()'s for-loop at that exact point, silently discarding EVERY crontab entry after line 120 (the 华山论剑 sect-tournament scheduler, the 牛人三部曲 PK-event scheduler, and anything else later in the 140-line file) on every single boot, not merely failing to spawn one random equipment drop in one wizard's personal room.

Fix: added if (!objectp(ob)) return 0; immediately after the TEMPLATE_D->create_object() call in equipmentd.lpc's create_dynamic(), restoring the function's own already-documented "returns 0 on failure" contract (mirrors the identical early-return guard already present at the top of the same function for ilvl < 1). Also fixed the caller-side chain in cangku.lpc: stored the return in the already-available ob local and guarded the ->move() call with if (objectp(ob)). Verified live: debug.log goes from this error firing deterministically on every single fresh boot (both via a first connection reaching logind.lpc's TIME_D lazy-load path, and via a direct wizard update /u/redl/cangku.lpc/update /adm/daemons/ equipmentd.lpc) to zero occurrences across multiple clean reboots post-fix. Filed upstream as PR #4. Added as a new confirmed instance of AGENTS.md §7.147, with a new note about the extra severity when an unguarded factory chain is reached from inside a bootstrap/scheduler loop rather than ordinary gameplay.

Standing cross-cutting patterns checked, confirmed clean

• §7.121 / §7.124 (float-into-int economy math / fraction-vs- percentage threshold): grepped the whole tree for float-literal threshold assignments and wimpy-style fields; all ~140 set("env/ wimpy", N, ...) call sites use consistent integer 0–100 percentages, and the one genuine float rate family (quest/skybook damage-scaling code) is correctly declared float throughout. Clean.
• §8.3a (private/nomask command-dispatch demotion): already confirmed clean at onboarding time (the #ifndef __SENSIBLE_ MODIFIERS__ shim is dead on this driver) — re-confirmed no new private nomask command_hook-shaped declarations exist.
• §7.112 (NPC init() unconditional call_out chain with no reconnect guard): the one remove_call_out()-then-call_out() pattern found (d/newbie/npc/qianbo.lpc's shopkeeper greeting) DOES carry the guard already, and is cosmetic (a greeting message) rather than the death/combat-state class this pattern targets. Clean.
• §7.123 (bare file-scope mapping/array literal killing a compile): the only 4 hits for the raw pattern are either legitimate multi- variable protected nosave mapping a = (...), b = (...), c = (...); declarations or inside /* ... */ comments. Clean.
• §7.126 (stale .c extension in saved door/exit data): no "__DIR__...\.c"-style saved references found anywhere. Clean.
• §7.129 (tell_room()/message() wrapper forwarding omitted exclude as literal 0): already fixed upstream in this exact repo before this project even started (see bugs-found item at the top of this file) — adm/kernel/simul_efun/message.lpc's message() correctly checks arrayp(exclude) || objectp(exclude) before forwarding to the real efun. Re-confirmed by reading the current source. Clean.
• §7.130 / §7.133 (net-dead / disconnect-notification gaps): net_dead() is defined on both clone/user/user.lpc and clone/ user/login.lpc, so the driver's real disconnect apply is handled. Clean.
• §7.131 (find_living/find_player without set_living_name): set_living_name IS called (7 sites) — this archive uses the modern registration convention, not the classic implicit-scan one. Clean.
• §7.132 (map()-over-mapping bound to the wrong argument): no map(m, (: $1 ... :))-shaped single-argument lambda calls over a mapping found anywhere. Clean.
• §7.134 / §7.135 (uninitialized accumulator/lazy-init guard inconsistency): the specific room_descs/add_my_desc() identifiers from the documented pattern don't exist in this codebase (different room-description architecture); this lib's own dbase accessor layer was already verified correctly guarded at onboarding time. Nothing newly found.
• §7.136 (command-soul strip with no race content to re-grant): N/A — this codebase's command dispatch (feature/command.lpc) is a different architecture (no cmdsoul_list/NPC_SOULS concept).
• §7.139 (interactive catch tell config flag / %^TAG%^ colour markers bypassing add_message()): this lib renders colour via direct ANSI escape macros (<ansi.h>'s HIC/HIY/NOR, etc.) embedded literally in output strings, not a catch_tell()-mediated %^TAG%^ translation layer — confirmed live throughout this session (every coloured message rendered correctly with real escape codes, no literal %^...%^ text ever appeared in any test transcript except one obscure, unreachable-in-practice quest macro in quest/ skybook/jimou/{fengbian,tianbian}.lpc that builds a %^WEATHER_ CHANNEL%^-tagged string with no corresponding tag-resolution table anywhere in the codebase — flagged as a content/quest observation, not fixed, since deciding what that tag should resolve to would be a content decision). §7.139's actual failure mode does not apply here.
• §7.141 / §7.142 (replace_program() fold breaking post-boot communication / virtual-object engine masking a broken exit typo): no virtual-object engine in use in this codebase; replace_program() calls are the standard ROOM-class pattern already covered by the corpus-wide §7.100 sweep (N/A here, this lib wasn't part of that sweep's target shape and its own replace_program(ROOM) calls are the harmless single-inherit-fold idiom).
• §7.143 (shop NPC add_action/force_me command_giver mismatch): live-tested at d/newbie/zahuopu's 钱伯 (qianbo, F_DEALER mixin) — list and buy fire both correctly dispatched to the NPC's own do_list/do_buy (confirmed via the correct in-character "钱伯冷笑道：穷光蛋，一边呆着去！" no-money rejection message, not a silent no-op). Clean.
• §7.144 (one-shot set_name() setup guard defeating per-instance rename): no generic reusable NPC base class self-naming in its own setup() found in this codebase's NPC/F_MASTER inherits.
• §7.145 (broken pre-check wired as direct verb override instead of base-class hook): not found; shop verbs (buy/list) dispatch correctly per the §7.143 test above, with no shadowing override.
• §7.146 (stray / meant as /*): zero hits in any .lpc/.h file (only in ASCII-art map/doc files, which aren't compiled).
• §7.148 (parameter literally named nosave/static): zero hits.

Interactive verification (this pass)

Registered a fresh Chinese-named test character (qinfengnt, 秦风, admin pw fluffos, play pw Mud2026play) via a raw Python socket bridge script, one continuous session:

• washto 20 20 20 20 → landed at 世界之树 correctly, innate skills banner shown, score/i both fully rendered with correct values.
• Walked to /d/newbie/lianwuchang (练武场), bai wu bo (拜师) succeeded ("古村第二代弟子"), score afterward correctly showed 【门派】古村 / 【师承】武伯.
• learn wu bo unarmed 1 correctly raised unarmed skill; skills confirmed 基本拳脚 (unarmed) - 不堪一击 6/14.
• Walked to /d/newbie/houcun-road, kill ye tu (野兔) — full combat round-trip with real damage numbers both ways, correct kill message, correct XP/potential/江湖阅历 reward banner. Safe low-level wild target, no dedicated "safe sparring" NPC found reachable in the newbie zone (d/heizhao/npc/muren.lpc's stat-mirroring training dummy exists in the codebase but its containing zone, d/heizhao/, is unrelated/unreachable from the newbie village — a wild rabbit served as the safe low-risk combat test instead).
• quit → correctly triggered this lib's documented "new account must stay online 30 minutes" retention policy (already documented as intentional design on sibling hhsj); confirmed with y, save file actually removed, zero debug.log errors throughout.
• Reconnected as the SAME character before deletion actually confirmed the persistence-then-delete race wasn't an issue (character correctly showed 古村第二代传人/门派 古村/师承 武伯/20-20-20-20 stats on a fresh reconnect prior to the final quit y), satisfying the "reconnect after a wall-clock gap, confirm state persisted" step.
• Admin account (fluffos/Mud@2026wiz, already seeded) reconnect verified clean: lands at WIZARD_ROOM as documented, goto teleport and update (live-recompile) wizard commands both work correctly.
• Shop test: goto /d/newbie/zahuopu, list/buy fire against 钱伯 both dispatched correctly (see §7.143 note above).
• Test character's save file (qinfengnt) removed via the quit y flow above; no other throwaway saves created this session.

RSS stayed under 130MB throughout (well below this lineage's documented §7.110 OOM risk threshold) — no mass-restore operation was exercised this session, so that risk wasn't specifically stress-tested here.

Local run

cd libs/nt7
~/src/fluffos/build-debug/src/driver config.fluffos

Port: 40211.