zhyx (中华英雄) -- porting notes

Source: git clone https://github.com/MudRen/zhyx (commit 762908319b23b4ae910718daf524f91c70bf5817, cloned 2026-08-26). Per an earlier research pass this is the canonical/current version of the "中华英雄" ("Chinese Hero") codebase -- its own README.md self-identifies as "中华英雄MUD FluffOS v2017 版", names mud.ren 5000 as its own live, long-running production server ("长期稳定开放"), and has a much larger domain list (~90+ named locations under d/) than a superseded sibling repo (ZHLib, an older 2006 snapshot with a far smaller domain set) that an earlier pass already ruled out. Not a duplicate of anything already in this collection -- checked against the yanhuangwuhun/YHLib/yhmud cluster, which is thematically similar (same underlying ES II/yh2003 lineage, see below) but has its own distinct, much smaller domain list. Slug zhyx, number 931, port 40233 (40232 was picked first but collided live with a concurrent session's sunshadow onboarding; reassigned before first boot).

Raw archive: 16,786 files (~99MB), 15,561 .lpc files after conversion. Top-level layout: adm b backup binaries clone cmds config.cfg d data doc driver dump feature grant hate include inherit kungfu log quest questobj shadow status temp u -- the mudlib root is the repo root itself (no nested wrapper directory). driver/binaries are empty placeholders (.gitignore-only); a bundled MudOSxp.exe (an old MudOS Windows binary, mentioned in the repo's own README as an alternative, password-incompatible way to run this lib) and a stray FreeBSD ELF binary at u/yuchang/driver were both ignored, per standing project policy -- only the LPC mudlib source was converted and run through this project's own FluffOS build.

Lineage

adm/single/master.c's header comment is the exact same lineage chain already documented for yanhuangwuhun/yhyxs/yhwhckdm in this collection: "for ES II mudlib / original from Lil / rewritten by Annihilator (11/07/94) / modified by Xiang for XKX (12/15/95) / updated by Doing Lu for hell (2K) / Modify By Linux@lxtx for yh 2003.3". Same adm/single/{master,simul_efun} layout (not adm/obj/), same F_DBASE-based property storage (no bare set/query/delete simul_efuns -- confirmed by reading adm/simul_efun/*.lpc), same overall registration flow (English login id -> confirm -> Chinese surname -> Chinese given name -> admin password -> regular password -> 1-5 character-type selection -> gender -> world entry) as yanhuangwuhun. But independently, substantially diverged: a byte-count diff of this archive's master.c/chinese.c/named.c/ logind.c/securityd.c/message.c against yanhuangwuhun's converted copies runs from 46 differing lines (chinese.lpc) up to 2,236 (logind.lpc) -- confirming a real, old fork rather than a near-copy, so every catalog fix below was independently verified against this archive's own code, not blindly ported. The whole domain/quest/ kungfu-skill content tree is this codebase's own -- distinct from every other yh2003-family lib already in this collection.

Pipeline

1. Cloned directly into libs/zhyx/raw/zhyx/ (git-hosted source, no archive-extraction step needed -- .git stripped before copying in, per this project's convention for already-git-hosted sources). 2. scripts/convert_lib.sh libs/zhyx/raw/zhyx libs/zhyx/work GB18030: already_utf8=642 converted=16092 lossy=13 skipped_binary=39; renamed 15,561 files to .lpc; fixed 6,812 literal .c" references + 17 #include <....c> refs; converted 13 local angle-bracket includes to quotes; static->nosave touched 143 files (no collision counterexamples found on inspection, unlike some sibling libs' large false-positive runs). 3. Full-tree straggler sweep (per AGENTS.md §4.1, beyond what convert_lib.sh's extension-based scan covers): a plain Python UTF-8-decode walk of the whole work/ tree found 24 non-UTF8 files. Triaged each: - Genuine un-transcoded GBK text (fixed): clone/game/8_hlp, clone/game/21_hlp (the 8-张/21-点 card-minigame help text -- the exact same extensionless-help-text gap class documented for yanhuangwuhun/yhyxs/yhwhpublicfi, though NOT the same specific files/content here -- this codebase's own card games) and d/death/README.TXT (an unreferenced dev note, a captured chat log about the reincarnation quest chain). Converted with iconv -f GB18030 -t UTF-8. - Genuine binary, correctly left alone: MudOSxp.exe (bundled driver), u/yuchang/driver (a personal wizard's own stray FreeBSD ELF binary), kungfu/skill/huashan-quan/MFM1992 (opaque binary blob, zero references anywhere in the tree by that name -- orphaned dead data alongside its sibling song.lpc skill file), and the adm/etc/{hzk12,hzk14,hzk16,asc12,asc14,asc16} + adm/etc/language/{hzk12,hzk14,hzk16,asc12,asc14,asc16,font.lib, zhiku1..6} files -- genuine binary Chinese bitmap font tables (HZK-format), confirmed still real and in use via adm/daemons/text2pic_d.lpc/text2d.lpc/kungfu/class/ultra/ robot.lpc (an ASCII-art/text-to-picture rendering feature). - False-positive from file(1)'s heuristic (no fix needed): include/tomud.h/tomud1.h -- pure-ASCII macro definitions that legitimately embed raw control bytes (0x19/0x1A) as literal protocol-marker characters inside string literals (a GUI-client display-tag format), not encoding damage. 4. Uppercase-.c-extension stragglers (missed by convert_lib.sh's lowercase-only rename): 9 files. Checked encoding first (all already UTF-8, no transcoding needed) then checked reachability before deciding what to do with each: - 4 confirmed LIVE, fixed (d/death/{BAIHUXUE,XUECHI1,XUECHI2, JIMIESI}.C -> lowercase .lpc): all four are real rooms, reached via plain lowercase __DIR__"baihuxue"/"jimiesi"/"xuechi1"/ "xuechi2" exit-string references from d/death/{heisenlin, yanluodian}.lpc and from each other's own exits mappings -- unreachable on this case-sensitive filesystem/driver until renamed (the original site likely ran on a case-insensitive Windows filesystem, where this was always latent and invisible). Copied content byte-for-byte (already UTF-8) to the new lowercase .lpc name, removed the stale uppercase original. - 5 confirmed orphaned, left alone: d/zhuanshi/china-aojue/ PERFORM/{AOJUE,JIANQI}.C and .../EXERT/{POWERUP,SHIELD}.C -- the sibling file d/zhuanshi/china-aojue.lpc does reference lowercase china-aojue/perform/+china-aojue/exert/ paths, but the *directory itself* is still named uppercase PERFORM/EXERT on disk (even its one already-.lpc file, PERFORM/ zhangzhen.lpc, is unreachable through that reference), and the game's actual "china-aojue" skill resolves purely through SKILL_D(x) (#define SKILL_D(x) ("/kungfu/skill/" + x)) to the entirely separate, already-correct kungfu/skill/china-aojue/ tree -- so this whole d/zhuanshi/china-aojue/ subtree (including its already-lowercased file) is unreachable dead content, not fixed. u/yuchang/zhuangbei/obj/weapon/stick/BB.C -- zero references anywhere by any case; a personal wizard's orphaned draft weapon item.

Fixes applied (driver-incompatibility bugs)

• §9-class, highest-impact finding this session: log_error()'s warning-vs-error string check used the wrong case. adm/single/ master.lpc's log_error(string file, string message) -- the apply the driver calls with EVERY compiler diagnostic from EVERY lazy compile -- classifies a diagnostic as a real error (vs. a benign warning to just log quietly) with if (strsrch(message, "Warning") == -1). This driver's own compiler emits lowercase "warning: Unused local variable '...'" (never capitalized "Warning"), so that check was always true -- literally every compile warning from every lazily-loaded room/NPC/item (an extremely common, mostly harmless diagnostic -- dozens appear in a normal boot log) was broadcast straight to whichever player happened to be connected at that moment: the raw config.fluffos default error message text for a normal player, or "编译时段错误：" + message for a wizard. Confirmed live and extremely reproducible before the fix: a single fresh character's very first room entry (many first-time lazy compiles firing at once) showed the default-error-message text up to 19 times in a single short session -- purely cosmetic (never blocked registration, movement, or any command), but a glaring, constant, unexplained-looking wall of "你发现事情不大对了， 但是又说不上来。" for every single real player, forever, until this fix. Root-caused via a temporary diagnostic write at the very top of master::error_handler() (a *different*, correctly-implemented apply for actual runtime error()s) that proved these specific messages were NOT going through that apply at all -- ruling out every runtime-error hypothesis and pointing at the separate compile-time log_error() path instead. Fixed by matching both cases:

``lpc if (strsrch(message, "Warning") == -1 && strsrch(message, "warning") == -1) `` Verified with 4 independent fresh-driver registration runs after the fix: zero spurious messages in any of them (vs. every single pre-fix run showing several).

• §15h (is_chinese GBK byte-range check) -- adm/simul_efun/ chinese.lpc's is_chinese() was the classic whole-string GBK lead/trail-byte-range check (str[i] < 161, an i%2==0 lead-byte gate, str[i] < 176 || str[i] >= 248). This driver's strings are UTF-8 codepoint-indexed, so every real Chinese character failed the check. Rewritten as a CJK Unified Ideographs codepoint range check (str[i] >= 0x4e00 && str[i] <= 0x9fff), dropped the byte-lead gate, floor changed < 2 -> < 1.
• §15h (check_legal_name length bound) -- adm/daemons/logind.lpc's check_legal_name(name, maxlen) had strlen(name) < 2 || strlen(name) > maxlen (maxlen=4 at both call sites, a byte-oriented "2 Chinese characters" GBK bound). Fixed to strlen(name) < 1 || strlen(name) > maxlen / 2. Also fixed the separate combined surname+given-name length check in get_name(): strlen(fname) < 4 -> < 2.
• §15h (named.lpc full "deep" fix) -- adm/daemons/named.lpc has the complete shape documented for this lineage:

- #define PATH(name) (name[0..1] + "/" + name) -> name[0..0]. - Every strlen(name) < 2 guard (remove_name, map_name, who_is, invalid_new_name) -> < 1. - A name[0..1] bucket-key reference in remove_name -> name[0..0]. - invalid_new_name()'s sliding-window near-duplicate-name check: guard strlen(name) < 4 -> < 2; loop bound i <= l - 4 -> i <= l - 2; 2-char window name[i..i+3] -> name[i..i+1]; 3-char window name[i..i+5] (guarded i+6<=l) -> name[i..i+2] (guarded i+3<=l).

• New variant of the message()/tell_room() 4th-argument bug -- adm/simul_efun/message.lpc's local message(mixed, string, mixed, mixed) wrapper called efun::message(arg, message, target, exclude) unconditionally; tell_room()'s call site (message("tell_room", str, ob, exclude)) passes a bare int 0 for exclude on every ordinary call (no exclude list), which fails efun::message()'s 4th-arg type check (void|object|object*). First reproduced live via securityd/preload daemons that call CHANNEL_D->do_channel() -> tell_room() during boot. Fixed by guarding: only pass the 4th arg when it's actually truthy:

``lpc if (exclude) efun::message(arg, message, target, exclude); else efun::message(arg, message, target); ` Unlike yanhuangwuhun's instance of this same bug, no separate forward-declaration workaround was needed here -- tell_room()'s bare message(...) call already correctly bound to this file's own override (confirmed by the crash's own stack trace pointing inside the local message() function body, not at tell_room()` directly).

• Extension-rename-width bug (.c->.lpc, same shape as the nitan_ceshi/nitan_san/yanhuangwuhun precedent) -- adm/daemons/ eventd.lpc and adm/daemons/storyd.lpc both do get_dir(DIR + "*.lpc") then map_array(names, (: $1[0..<3] :)) to strip what used to be the 2-character .c extension, now leaving a stray .l ("emei.lpc"[0..<3] -> "emei.l", not "emei") -- surfaced as call_other() couldn't find object '/adm/daemons/event/emei.l' during eventd's preload. adm/npc/luban.lpc has the identical idiom at 2 further sites (a player-built-house NPC clearing stale dbase entries by filename). All 4 sites widened [0..<3] -> [0..<5].
• UDP DATAGRAM socket read_callback receives a BUFFER, not a string -- adm/daemons/network/dns_master.lpc (DNS_MASTER, on the real preload list and actively listening on a real UDP port) declares void read_callback(int sock, string msg, string addr), but this driver's socket_create(DATAGRAM, ...) delivers the packet as a buffer, not a string. Crashed sscanf() with "Bad argument 1 to sscanf Expected: string Got: BUFFER" on every real incoming packet -- and this daemon DOES receive real incoming packets: the crash was first caught live from an actual Intermud-style "mudlist ping" UDP packet sent by another real, currently-running Chinese MUD ("YHMUD"/"炎黄", HOST:...||PORT:5566||...) somewhere on the public internet, arriving unprompted during this port's boot-test window. Fixed by widening the parameter to mixed and coercing defensively (this driver has no to_string() efun; string_decode(buffer, encoding) is the buffer->string conversion it actually provides):

``lpc if (bufferp(msg)) msg = string_decode(msg, "UTF8"); ` Flagging per AGENTS.md/task convention: this is a genuine, active outbound/listening network side effect, unlike almost every other lib in this collection (all sandboxed, no real connections). This lib's dns_master.lpc binds a real UDP socket and both sends and receives real traffic to/from real hosts on the public Intermud network at every boot. Should NOT be swept into high-frequency automated reboot loops (§10.0-style long-sit scans, round-two/ round-three re-test cron) the way every other lib in this collection safely can be -- each boot is a real network event against real third parties. The other 3 files under adm/daemons/network/ (ms.lpc, mail_serv.lpc, netmail.lpc -- an Intermud-style cross-mud mail relay) are confirmed genuinely dead/unreachable: not on the preload list, and cmds/usr/mail.lpc (the real, live in-game mail command) uses a completely separate, unrelated MAIL_D (/adm/daemons/ maild) system -- left as-is, several undefined-macro compile failures in that dead subtree (MAILDIR/POSTAL_D/MAILER_D/ THIS_MUD/SAVE_EXTENSION/MUD_TYPE`) not fixed.

• (int) cast is compile-time-only, not a real runtime coercion, on this driver (see this project's own standing reference note) -- hit this TWICE independently in this lib:

- adm/daemons/versiond.lpc's in_server(): port = get_config(__MUD_PORT__) + VERSION_PORT; crashed socket_bind() with "Bad argument 2 ... Expected: int Got: \"12\"" (i.e. the string-typed get_config() result silently string-concatenated with the bare int VERSION_PORT, 12, rather than adding). A first attempt adding a plain (int) cast did NOT fix it (same crash, same value) -- switched to to_int(), which actually coerces: ``lpc port = to_int(get_config(__MUD_PORT__)) + VERSION_PORT; ` (Left the resulting bind failure -- this driver's config-array index for the mud's own port evidently isn't retrievable the way this old code expects -- to fail gracefully through the function's own existing if (socket_bind(...) < 0) { log(...); return; } guard, which now works correctly and logs cleanly instead of crashing. VERSION_D's own version-sync feature may not be fully functional as a result, but it's a narrow admin/release-tracking daemon, not on the registration/gameplay path.) - include/getconfig.h's shared LOCAL_PORT() macro had the exact same ((int) get_config(__MUD_PORT__)) shape, hit by adm/ daemons/network/messaged.lpc's startup_udp() (lazily loaded the first time ANY character successfully logs in, via logind.lpc's check_ok()) -- same crash shape, same root cause. Fixed once at the shared macro definition (covers this call site and d/city/ npc/shizhe.lpc's use of the same macro): `lpc #define LOCAL_PORT() (to_int(get_config(__MUD_PORT__))) ``

• report_error() declared void but its result used as a string argument -- adm/single/master.lpc's error_handler() (the real runtime-error apply, separate from log_error() above) does tell_object(this_player(1), report_error(error)) for any non-wizard player, but report_error(mapping error) was declared void and never returned anything (it only had the side effect of posting to the "wiz" channel) -- so tell_object()'s 2nd argument evaluated to 0, not a string, crashing tell_object() itself for every non-wizard player who ever triggers ANY genuine runtime error site-wide, and -- because the crash happens partway through error_handler()'s own body -- silently preventing that function's own final write_file(LOG_DIR + "error_handler", ...) line from ever running, so these secondary crashes left zero trace in any log. Fixed by making report_error() build and return the message string (matching its own declared use), and wrapping its CHANNEL_D-> do_channel(this_object(), "wiz", ...) side-effect call in catch() since this_object() here is master itself, not a normal player/wizard speaker, and the channel code path isn't designed for that case.
• Missing macro, §8g-style (target file exists, macro just never defined):

- SKI_ZHEN -- inherit/skill/array.lpc (a "formation"/阵法 skill base class) has its own header comment literally saying * #define SKI_ZHEN, but the real #define was never added anywhere. 4 files (kungfu/skill/{quanzhen-jianzhen,zhenwu-jianzhen, dagou-zhen,hebi-jianzhen}.lpc) inherit SKI_ZHEN;. Added #define SKI_ZHEN "/inherit/skill/array" next to SKILL/ SHAOLIN_SKILL in include/globals.h. - IP_D -- adm/daemons/ipd.lpc exists and defines exactly the one function its only call site needs (seek_ip_address(), from cmds/wiz/ipname.lpc), but the macro was never defined. Added #define IP_D "/adm/daemons/ipd".

• Typos, individually diagnosed (not part of any wider bug class):

- u/kasumi/kungfu/snake.lpc -- a fullwidth ！ (U+FF01) placed *outside* a string's closing quote as bare code syntax (same §9 class as the mamian.lpc/） precedent in yanhuangwuhun). Moved inside the string, where the sentence clearly intended it. - u/sanben/yuanxiao/new_lot_machine.lpc -- #define SECOND_PRIZE /d/death/obj/tianjing" (missing the opening quote). Added it. - inherit/condition/damo.lpc -- the same shared-root bug already documented in yanhuangwuhun's NOTES.md, present here byte-for-byte identically: an unclosed string in die_reason() (return name + HIM "发作身亡了; missing its closing ") plus a separate bracket-mismatch typo in update_condition() (intp(cnd["remain"] || ... missing a )). Fixed both exactly as in the sibling lib.

Confirmed NOT needed (checked, not applicable)

• §4/§7.1 (master's lazy security-daemon load recursion) -- valid_read/valid_write in master.lpc only call find_object(SECURITY_D), never load_object(); valid_write degrades to deny (return 0) and valid_read degrades to allow (return 1) when securityd isn't loaded yet -- no recursion risk, confirmed by reading both functions in full.
• §7.2 (missing get_root_uid/get_bb_uid) -- both already implemented in master.lpc.
• §7.3 (create() destructing SIMUL_EFUN_OB) -- master::create() only does one write(); no destruct() anywhere in it.
• §15n (custom securityd ACL blocking compile-time access) -- adm/daemons/securityd.lpc does have the same exclude_read/ trusted_read default-deny-by-directory shape as other libs in this family, but its early-allow list already covers restore_object/ file_size, and the actual boot/registration/gameplay testing below never once hit an ACL-denial error for load_object/ recompile_object/include -- the ROOT_UID/(admin) early-allow earlier in the function already covers the compile-time access path in practice for this lib. Not patched (no evidence it's needed; a (admin)-status wizard test also confirmed no denials).

Boot / registration verification

• ~/src/fluffos/build-debug/src/lpcc, single-file, confirmed clean on all core base classes before ever running a full batch sweep (per this project's RAM-safety discipline): adm/single/master, adm/single/simul_efun, and a scoped --batch run over all 174 feature/*.lpc + inherit/**/*.lpc files -- 173/174 passed on the first attempt, the one failure (inherit/condition/damo.lpc, a content bug, see below) fixed and reconfirmed 174/174 before proceeding.
• Full lpcc_check.sh batch sweep (all 15,565 objects): peaked at ~2.7GB RSS (well under the 10GB safety threshold), completed in ~2.5 minutes. 15,354 / 15,565 = 98.6% pass (211 failures) on the final sweep, after also fixing inherit/condition/damo.lpc, u/kasumi/kungfu/snake.lpc, and u/sanben/yuanxiao/ new_lot_machine.lpc (see below); remaining failures triaged by category -- the large majority trace to a personal wizard sandbox (u/yuchang/zhuangbei/, an entire parallel weapon/armor item collection whose own private weapon.h/armor.h define macros like STICK/SPEAR/MACE/ RAKE/FINGER pointing at a /std/weapon/...//std/armor/... path that doesn't exist anywhere in this archive -- a genuine, pre-existing content gap, not fabricated a substitute base class, per standing policy) plus a handful of narrow, genuinely unimplemented content gaps (attack_power()/defense_power()/ damage_power() used but never defined anywhere, affecting only the Tang-Sect "钵孟锁物"/"唐门暗器" hidden-weapon throwing skills; check_owner() used but its base class feature/oneowner.lpc never inherited, confined to one wizard's personal "宝石"/gem item drafts; a few more §8g-style dead-subtree undefined macros already covered above). None of the 218 affect the boot/registration/gameplay path actually exercised below.
• Real build-debug driver boot on port 40233: clean, zero fatal debug.log/log/error_handler errors (only benign compile warnings, correctly routed to log/log_error after the fix above, never shown to a connecting player).
• Full registration flow verified via raw socket scripts (scripts/ mudclient.py), 5 independent runs, real Chinese names each time (id/surname/given-name: qinling/秦/岭, qinhuai/秦/淮, qinchuan/秦/川, qinzhen/秦/真, qinshui/秦/水), each reaching the actual starting room "世外桃源" ("Peach Blossom Land," a scenic entry area with 2 present NPCs, 水笙/狄云), look correctly re-displaying it, score correctly returning the graceful "还没有出生呐，察看什么？" (not-yet-born) response rather than erroring, and quit cleanly returning "欢迎下次再来！". The last 2 of these 5 runs (post the log_error() fix) showed zero spurious error-message noise; all 3 runs before that fix showed it repeatedly (3 to 19 times per session) -- see the fix write-up above.
• Admin account seeded: id fluffos, 普通密码(login) Mud@2026, 管理密码(recovery) Admin2026, display name 浮浮 (surname prompt skipped with a bare newline, given name 浮浮 -- a single 浮 alone is rejected by the 2-character-minimum combined-name check; this lineage's combined-length floor is intentional content, not a bug). Granted (admin) via fluffos (admin) appended to adm/etc/ wizlist (the same plain-text wiz_name status format already used for the existing mudren (admin) entry). Verified: fluffos/ Mud@2026 login succeeds, look shows the wizard-only room-path/ coordinate suffix (/d/register/entry - (0,0,0)), confirming (admin) status is live; quit clean.
• Save files for the orchestrator to add (not gitignored, normal add): libs/zhyx/work/data/user/f/fluffos.o, libs/zhyx/work/data/login/f/fluffos.o. All other test characters' save data (6 throwaway ids) was deleted before finishing, per convention.
• Outbound network check (per task step 6): the one intentional exception in this whole collection so far to "every lib is fully sandboxed" -- see the dns_master.lpc UDP fix write-up above. Confirmed via direct observation of a real incoming packet from another live MUD during a routine boot-test window, not a hypothetical.

Not run this session

• LPC formatter (§9): node is not available in this environment (command not found: node), so tools/lpc-syntax/bin/format-corpus.mjs could not be run against the handful of files edited above. Every edit in this session was small (a few lines each, one macro definition, one string-literal fix, one cast-vs-conversion swap) and was manually reviewed against the surrounding code's existing indentation/brace style rather than auto-formatted. Worth a formatter pass in a future session that has node available.
• Full deep §10.7-style playthrough / WASM pass: out of scope for this onboarding session per the task brief (native boot + registration verification only). dns_master.lpc's live network side effect is an extra reason to be deliberate about how often this lib gets rebooted in any future automated long-sit/WASM-enablement pass.

深度功能测试（round two, AGENTS.md §10.7）— 2026-08-27

First live *playthrough* pass (prior session only verified registration + look/score/quit/admin-install, per the sections above). Read doc/help/newbie and doc/help/yxt_newbie first — the intended path is register → walk to one of 4 personality NPCs off 世外桃源 (東/西/南/北) → out to commit → 阎罗殿 → wash (reroll stats) → born <中文地名> → real starting room; fight is the built-in safe-sparring command ("点到为止，因此只会消耗体力，不会真的受伤"); teacher/sect join is bai <NPC>, gated on the NPC's own willingness ("很多师父不轻意收徒"); new-player stat gift is ask zhyx_boss about 天赋 + add <stat>. Played a continuous session as two fresh characters (qintestb, female, kept as the representative test character; qintesta, used to find/verify the bugs below, then deleted) on the native driver (build-debug), plus targeted admin-eval probes for the death/PvP path that a level-1, mostly-miss-everything character can't reliably reach live in reasonable real time. Booted/rebooted the driver 6 times across this pass (log/error_handler/log/log_error are append-only across restarts, not truncated — a wc -l diff plus checking a process's own start time via stat/ps was needed more than once to tell a genuinely-new crash from a stale leftover entry near the tail).

Bug 1 (new class, AGENTS.md §7.125): the email-registration gate is permanently dead from the very first character ever created

adm/daemons/logind.lpc's enter_world() — which runs on EVERY login, not just character creation — does an unconditional user->set("registered", 1); //user->set("born",1); a few lines before its OWN later if (!user->query("registered")) ... checks, right after the unrelated starting-clothing block. This permanently defeats: d/register/entry.lpc's valid_leave() ("you must register <email> before leaving 世外桃源") and adm/daemons/channeld.lpc's "you must register before using channels" gate. Reproduced live: a brand-new qintesta walked east out of the starting room with zero email ever registered — no rejection, no crash, just silent success. Root-caused by temporarily deleting the statement and confirming a fresh character was then correctly blocked ("你还不快注册？") until register <email> + decide were actually run against 水笙, then correctly let through afterward. Fix: deleted the premature user->set("registered", 1); from enter_world() (adm/daemons/logind.lpc:959) — the real, correct place this flag gets set, d/register/npc/shuisheng.lpc's do_decide(), was already there and always correct, just permanently short-circuited. Verified with a second full fresh registration (qintestb) post-fix: blocked from leaving pre-registration, walks free immediately after register+decide. See AGENTS.md §7.125 for the generalized writeup and the sibling libs worth checking (yanhuangwuhun/yhwhckdm/yhyxs/yhwhpublicfi, same lineage, likely same enter_world() shape).

Bug 2 (extends AGENTS.md §7.12): the onboarding-time message()/tell_room() fix was silently dead code for the wrapper's own internal callers

The onboarding session already added the standard if (exclude) efun::message(...) else efun::message(...) guard inside adm/simul_efun/message.lpc's local message() override, and it was verified working for its own original repro (CHANNEL_D->do_channel() at boot). This pass hit the *exact same* *Bad argument 4 to EFUN message() Expected: object, array, Got: int(0) crash again, live, reproducibly, from plain ambient gameplay (clone/misc/corpse.lpc's decay() calling a bare 2-arg tell_room(env, msg) on its own 60s/30s call_out timer — no player action needed at all). Root cause, isolated via three direct eval tests: on this driver, an unqualified call to an identifier that is ALSO a genuine hard efun (message is a real FluffOS efun) resolves to the hard efun, not to a same-named local function, specifically when the call is made from within the very file/object that defines that local function — tell_room() lives in the SAME FILE as the guarded message(), so its own internal message("tell_room", str, ob, exclude) statement was binding straight to the hard efun the whole time, completely bypassing the "already fixed" guard sitting a few dozen lines below it. Every EXTERNAL caller of message(...)/tell_room(...) elsewhere in the mudlib was and is correctly routed through the guard (confirmed: an identical bare call from a plain eval, i.e. a different object, worked fine) — this trap is narrowly about the wrapper file's own internal self-calls. message_system() and shout() in the same file had the identical bare-self-call shape with their own literal/falsy 4th argument (all_interactive(), 0 and this_player() respectively, the latter frequently 0 in the NPC-death-event contexts several NPCs actually call shout() from) and got the same fix pre-emptively, even though only tell_room()'s instance was live-reproduced. Fix: qualified all three internal self-calls with efun:: explicitly (verified this correctly bypasses the same-object shadowing, alongside a call_other find_object(SIMUL_EFUN)->message(...) which also worked), keeping the exact same exclude-guard logic:

varargs void tell_room(mixed ob, string str, object *exclude)
{
    if (! ob) return;
    if (exclude)
        efun::message("tell_room", str, ob, exclude);
    else
        efun::message("tell_room", str, ob);
}

Verified live post-fix on a completely fresh driver boot (to rule out stale-log false positives from a still-shutting-down prior driver process, which briefly looked like the fix had failed before the timing was untangled): tell_room(), message_system(), and shout() all now execute cleanly via direct eval, and zero new log/error_handler entries for this signature appeared across a full fresh boot + registration + combat + shop + gift session afterward. See AGENTS.md §7.12's new addendum for the general lesson (verify a message()/tell_room() fix by testing the shared wrapper's OWN internal helpers directly, not just the originally-reported repro path) — worth a quick eval tell_room(some_room, "test\n") sanity check on any other lib that received this project's standard §7.12 fix.

Bug 3 (extends AGENTS.md §7.11): live PvP kills silently corrupt killer/victim combat-tracking state forever

adm/daemons/combatd.lpc::killer_reward()'s player-vs-player branch calls log_file("nosave/killrecord", ...) — a directory (/log/nosave/) this archive's shipped work/ tree never contains — immediately BEFORE killer->remove_killer(victim); in the same if block. Confirmed via a controlled admin-triggered kill (eval calling ob->die(ob) against the test character, since no level-1-reachable live opponent could realistically kill a fresh character in test-session time) that this throws uncaught (*Wrong permissions for opening file /log/nosave/killrecord for append. "No such file or directory"), aborting the rest of the block — remove_killer() never runs, so every real PvP kill leaves the killer's killer/want_kills tracking on the victim and the victim in the killer's enemy list forever. This is the same log_file()/assure_file() gap already cataloged extensively in AGENTS.md §7.11 (adm/simul_efun/file.lpc's log_file() was a bare write_file(LOG_DIR + file, text) with its own correct, already-defined assure_file() helper sitting unused two functions below it) — same family of combatd.lpc::killer_reward() instance already documented for jyqxc2013fwq (that one broke the death/resurrection transition itself, worse blast radius; this one corrupts combat state instead, since this lib's own death-room move happens earlier in die(), unaffected). Fixed at the shared wrapper rather than only the one call site: added assure_file(LOG_DIR + file); inside log_file() itself (plus the one-line forward declaration this driver needs since assure_file() is defined textually after log_file()), closing the identical latent gap for every other log_file() call into a never-shipped subdirectory project-wide in one fix — grepping log_file("<dir>/... call sites against find work/log -type d turned up at least job/ (雪山 burning quest logging), test/ (神龙教 PK/force-join records), and voting/ (polld.lpc) as other genuinely-referenced-but-never-shipped subdirectories that this same fix now protects, not just nosave/. Verified live on a completely fresh boot with NONE of these directories pre-existing on disk: log_file("nosave/killrecord", ...) and log_file("job/xueshan", ...) both auto-created their target directories on first use and succeeded with zero error. (Note: log/ is entirely gitignored project-wide as driver-recreated runtime state — libs/*/work/log/ in the top-level .gitignore — so this fix deliberately lives in code, not as a committed placeholder directory; a bare mkdir workaround tried earlier in this session would not have survived a fresh checkout.)

Confirmed clean (checked, cross-cutting bug classes from other libs' deep-test passes)

• §7.121-style currency float corruption: not applicable. This lib's economy (feature/banker.lpc's do_check/do_convert/ do_deposit/do_withdraw, clone/money/*.lpc's base_value) is pure integer denomination math (base_value values are plain literals like 1/100/10000) — no exchange-rate float anywhere in the currency path.
• §8.3a private command_hook: feature/command.lpc (the one actually inherited into the player body via the COMMAND macro) correctly declares protected nomask int command_hook(string arg). A second, unrelated feature/command1.lpc DOES have the exact buggy private nomask int command_hook(...) shape, but it's dead code — zero files inherit it anywhere in the tree (grepped for COMMAND1/command1" and found nothing) — left alone as an inert archive artifact, not a live bug.
• §7.122-style class-item duplication on reconnect: no auto_load/compute_autoload_array/TMI-2-style mechanism exists in this codebase at all (grepped, zero hits) — not applicable in the general sense. Did specifically check the one candidate that looked suspicious on paper: enter_world() unconditionally clones and move()s a /clone/misc/fly "新手飞行包" into inventory every login while age < 16 (no present() guard visible in the source). Tested live across a real quit + reconnect cycle expecting to find duplicates — found exactly ONE copy both times, no accumulation. Did not fully root-cause why (a no_drop/no_get/no_put item like this may not round-trip through ordinary save_object()/restore_object() the way ordinary inventory does, so each login's fresh clone may be the only copy that ever really persists), but the live test is unambiguous: this is NOT a reproducible duplication bug on this lib, despite superficially matching the §7.122 shape. Documented here so a future pass doesn't waste time re-suspecting it without re-testing.

What else was tested and confirmed working

• Registration → personality → born, twice, full flow, both genders (qintesta male "光明磊落", qintestb female): id/surname/given-name → 2 passwords → role-type → gender → 世外桃源 → register+decide → walk to a personality NPC (陆天抒, east) → out → 阎罗殿 → wash → born 扬州人氏 → landed in 武庙 with starting cloth/shoes/fly bag, look/score/i all correct at every step.
• Combat (safe-sparring): fight <NPC> against 流氓头 (a weak street NPC near 中央广场) ran several dozen non-lethal exchanges, correctly self-terminated once 气/精 dropped low (no halt needed), awarded 潜能 without ever risking real death — matches the newbie doc's own description of fight exactly.
• Skill/sect acquisition: organic route tested via bai <丐帮一袋 弟子> (found via 中央广场's enter dong → 树洞内部, matching the newbie doc's "丐帮在扬州城里，易于寻找" note) — correctly returned a pending "对方还没有答应" state rather than an immediate accept/crash, consistent with the doc's own "很多师父不轻意收徒" description; did not force acceptance (would require deciding game-balance timing, out of scope). New-player gift route: ask zhyx_boss about 天赋 → add str correctly granted and applied one of the 6 free stat points, score reflected it immediately.
• Shop/economy: list/buy jitui at 醉仙楼 correctly rejected an impoverished character ("店小二冷笑道：穷光蛋，一边呆着去！") rather than crashing or silently succeeding — matches score's own "你目前 没有存款" for a level-1 character with no money yet; a real successful purchase was not reached live (would need actual gold, out of budget for this pass) but the rejection path itself is now confirmed clean post-Bug-3-fix (the same log_file() gap could plausibly have hit a shop/economy log path too, though none was actually reached live).
• Death/respawn: reached via a controlled admin eval (ob->die(ob) against the non-admin qintestb/qintesta test characters, never the admin's own body) rather than a live kill, since no level-1-reachable opponent could realistically kill a fresh character within this session's time budget — this is the path that surfaced Bug 3 above. Did not additionally verify the full reincarnate/revive-room cycle live (out of budget after the Bug 3 investigation); worth a follow-up pass with a genuinely lethal opponent or a longer session.
• quit → log/error_handler/log/log_error grep → reconnect after a real ~65s wall-clock gap: done twice (once pre-fix on qintesta, once post-all-fixes on qintestb). Both times: clean "这样东西不能 随意丢弃"(fly bag, expected — no_drop, harmless) + "欢迎下次再来！" on quit, zero new fatal signatures in either log immediately after, and a genuine reconnect after the wait landed back with correct persisted state (score/i matching pre-quit exactly).
• Outbound network / long-sit caveat: not re-tested this pass (out of scope for a §10.7 playthrough) — the onboarding session's dns_master.lpc UDP finding stands; this lib should still not be swept into high-frequency automated reboot loops.

Minor observation, not fixed (deployment/logging quirk, not a clear LPC bug)

adm/daemons/cleard.lpc's auto_clear() (a 30-minute recurring call_out) does cp("/log/debug.log", "/driver/system.log") unconditionally, and log/debug.log does not exist in this project's own way of running the driver (this project redirects the driver's stdout to an external file via its own `nohup .../driver config.fluffos

some.log` convention, rather than relying on the mudlib's internal

debug log file config value ever actually being written to inside the mudlib tree) — so this cp() throws "lstat failed" every 30 minutes, harmlessly (nothing else in auto_clear() depends on its result). Left as-is: fixing it would mean guessing whether this project's own driver-invocation convention should change, or whether the mudlib should stop assuming a log/debug.log inside its own tree, either of which is a judgment call beyond "make the code's own already- intended logic work," not a clear-cut programming bug.

Test character / evidence

qintestb (female, 光明磊落, 扬州人氏) kept as the representative playthrough character — registered, born, has the new-player gift's 6 free stat points applied (add str used), visited 醉仙楼, survived a real quit+reconnect cycle. Save files: work/data/user/q/qintestb.o, work/data/login/q/qintestb.o. qintesta (used to find/verify Bugs 1-3) deleted after use, per convention. Admin account (fluffos/ Mud@2026) untouched, still seeded from the onboarding session.

Not run this session (round two)

• LPC formatter (§9): node still not available in this environment (command not found: node). All edits this round (adm/daemons/logind.lpc, adm/simul_efun/message.lpc, adm/simul_efun/file.lpc) were small, manually reviewed against surrounding style. Worth a formatter pass in a future session with node available.
• A live (non-admin-triggered) death/reincarnate/revive cycle and a real gold-funded shop purchase: both would need either a much longer session or a deliberately-outmatched fight; documented above as unverified-live rather than silently presented as tested.

wasm_status 审计（2026-09-01）：内容本身干净，closed.lpc 一处 VERSION_D 守卫补齐，验证为 playable

meta.json 的 wasm_status 此前一直留空（此前只做过原生驱动验证）。 本次批量审计（见 [[project_wasm_status_audit]]）补上这一步。

好消息：当前生效的 adm/daemons/logind.lpc 里对 VERSION_D->is_version_ok() 的调用早就被注释掉了（同目录下还留着 一份没有接入配置的 zhyx_logind.lpc，那份还是活的调用，但不是 LOGIN_D 宏指向的文件，不影响实际登入），questd.lpc 里的同类调用 也已经注释——说明这个登入路径本身在 WASM 下不会被同宗 revive/zhonghua2 家族那个"VERSION_D 编译失败→登入被拒"的闸门 卡住。只有 adm/daemons/closed.lpc 的 heart_beat()（开机常驻） 还留着一处未加保护的 VERSION_D->is_version_ok() 调用，同 AGENTS.md §1.3(c) 目录里的既有模式，补上 find_object(VERSION_D) && 前缀， 消除每次心跳都要重新编译一遍失败的 versiond.lpc 的额外开销。

system/kernel 之类没有独立 simul_efun 级 socket 依赖，dns_master. lpc/ftpd.lpc/versiond.lpc/messaged.lpc 等网络精灵编译失败 只是预载列表里的正常"外围精灵跳过"，不影响真实登入。

完整会话验证：英文 ID（3-10 个英文字母）→ 中文姓氏（留空）→ 中文 名字 → 双重密码 → 角色资质(3) → 性别(m) → 落地"世外桃源"，水笙/ 狄云在场，水笙主动提示注册邮箱 → look 正常重复显示房间描述 → score 得到"还没有出生呐，察看什么？"（与 NOTES.md 此前记载的 "score 显示未出生是正常游戏设计"一致，不是 bug）→ quit 干净退出 （"欢迎下次再来！"），全程无未捕获错误。wasm_status 设为 playable。

深度功能测试（2026-09-04，round three，shop + 拜师）

新角度：2026-08-27 那一轮醉仙楼 buy jitui 只走到「穷光蛋」拒绝 路径，丐帮一袋弟子 bai 也只挂起「对方还没有答应」。本轮补成 功购买 + 真正收徒。

实测过程

管理员 fluffos / Mud@2026（管理密码 Admin2026，本轮用普通 密码进）。账号还没 born，落地世外桃源，score 仍是「还没有出 生呐」——内容规则。巫师 goto 不受影响。

cmds/wiz/clone.lpc 额外要求 me->is_admin()，而 clone/user/user.lpc 的 is_admin() 只认 uid mudren（线上站 点锁，passwd 设 admin_flag 的路径已注释掉）。clone /clone/money/gold 因此回「你没有使用该命令的权限。」未改——这 不是崩溃，是这份档案自己的管理权限模型。改用 eval new("/clone/money/gold")->move(me)（cmds/adm/eval.lpc 只查 wizardp()）把金锭放进物品栏。

goto /d/city/zuixianlou，list 烤鸡腿八十文铜板。buy jitui 成功（「你从店小二那里买下了一根烤鸡腿」），找零九十九两白银 + 二十文铜钱（10000−80=9920）。本档案 F_DEALER 没有 XKX 那种丐 帮穷叫化拒绝，已拜师也能买。

goto /d/gaibang/inhole。左全的源码在 kungfu/class/gaibang/zuo-qu.lpc，但这份地图没把他放进房间（树 洞放的是黎生 + 随机一袋/二袋弟子）。一袋弟子不 inherit F_MASTER，所以 08-27 那次 bai 只会挂起。apprentice li 拜黎 生一次成功：permit_recruit() 允许无门派，score 虽然被未出生 挡住，但收徒文案是「恭喜您成为丐帮的第二十代弟子」，user.o 里 family_name":"丐帮" / master_name":"黎生" / 头衔「丐帮第二十 代传人」。save 有 30 秒节流，等过后再存成功。巫师不受 60 秒 last_on 重连冷却限制。

本轮没有新的 programming bug。cleard.lpc 每 30 分钟 cp("/log/debug.log") 仍会在 log/error_handler 留下 lstat failed（08-27 已记录，未改）。驱动 debug.log 本轮没落到 libs/zhyx/log/ 也没落到 work/log/，运行期错误看 work/log/error_handler（本轮游玩期间行数未增加）。