xkyx3b — 侠客英雄传 III

Archive: 侠客英雄传III 可用.zip. Port: 40036. Status: done (boots clean and quick, full registration flow verified end-to-end including a real Chinese name).

What this is

"侠客英雄传" — its own connection banner describes the setting as "以 金庸小说、西欧中世纪传说及漫画为背景的世界" (a world set against a backdrop of Jin Yong novels, Western European medieval legend, and manga — an unusual crossover premise). adm/obj/{master,simul_efun} layout. Small lib, ~3,839 raw files.

Fixes applied

1. AGENTS.md §15h, with a pre-existing (non-UTF8) typo along for the ride: adm/daemons/chinesed.c's is_chinese() had a for loop that always checked str[0] instead of str[i] — a genuine authoring bug predating this pass, not introduced by conversion. It was harmless in practice because the only caller, logind.lpc's check_legal_name(), always invokes it with a fresh per-character slice (CHINESE_D->is_chinese(name[i..<0])) — so "always check index 0 of whatever was passed" is equivalent to "check character i of the original name." Fixed to the standard single-character CJK codepoint check, which preserves this same effective (harmless) behavior while correcting the actual GBK byte-range logic (dropped the now- meaningless even-byte-length check and the loop, since only the first character of the argument ever mattered anyway). 2. check_legal_name()'s bound strlen(name) < 2 || > 12 → < 1 || > 6 (halved to character count), removed the i % 2==0 even-byte-offset gate. 3. AGENTS.md §15p: /adm/daemons/network/dns_master was in adm/etc/preload — removed proactively before the first boot attempt. Booted clean in well under 10 seconds, zero compile errors, no hang.

Interactive test result — full registration flow

No hidden pre-id prompts here (banner → stats → straight to the English id prompt). Verified the complete registration path in one continuous connection: id xkyxbtest → confirm y → real Chinese name 秦风 (avoiding "张三丰"/Zhang Sanfeng — a real historical/wuxia figure found in this lib's banned_name list, alongside game-specific terms) → accepted, proceeds straight to "请设定您的密码：".

lpcc sweep

3,178 files, 3,116 pass / 62 fail (98.0%). Failure tail is the usual shape (missing globals, a handful of syntax typos) — not triaged individually per AGENTS.md §6b/§13. Memory stayed healthy throughout (~13GB free).

Re-verification pass (2026-07-23)

Boot + registration re-tested end-to-end (real Chinese name 秦岭, English id + race selection this time, since the original pass only verified up to the password prompt). Found and fixed one new regression, same bug class as AGENTS.md §15ah:

• Missing /log/nosave/ directory (archive shipped without it, same "RAR/zip drops empty dirs" class as several sibling libs): enter_world() in adm/daemons/logind.lpc unconditionally log_file("nosave/WARNING", ...)s a potential/exp sanity-check note for every new player, and adm/simul_efun/file.lpc's log_file() did an unguarded write_file() with no directory to write into — this runtime error aborted the rest of enter_world(), so the new character was never actually move()d to the start room (look showed the literal empty void, "你的四周灰蒙蒙地一片"), and quit's message("system", ..., environment(me), me) then crashed on environment(me) being 0 (Bad argument 3 to EFUN message()). Fixed by creating work/log/nosave/ and hardening log_file() (wrap write_file() in catch()) and cat() (guard with stringp()) in adm/simul_efun/file.lpc, matching the established pattern from sibling libs. Re-verified with a fresh registration (id xkyxfixver, name 秦岭): look now shows the real start room content (the color-dog greeting from d/beginner/start1.lpc), score renders correctly, and quit completes cleanly with zero runtime errors in debug.log.

Driver rebuild / formatter / WASM pass (2026-07-23)

• LPC formatter (tools/lpc-syntax/format-corpus.mjs) run over all work/*.lpc: 3,178 total, 3,146 written, 21 already-idempotent, 11 refused (self-check errors — expected on messy legacy code, not chased).
• Native retest against the freshly-rebuilt driver (build-debug, rebuilt from latest upstream master today): clean boot, zero fatal errors. Full registration flow re-verified end-to-end in one continuous connection (id qinlingueb, real Chinese name 秦岭岳, race 0/human, gender m): look showed the real 小花狗/color-dog start-room greeting, score rendered the full character sheet, quit completed cleanly. Zero code changes needed — the reformatted source booted and played identically to before formatting.
• WASM test (scripts/wasm_client.js against build-wasm): boots clean through preload (only the expected non-fatal ftpd preload error, same shape as native). The banner displays correctly, but the single WASM connection then dies during logind.lpc's visitor-counter bookkeeping (write_file(LOG_DIR "login/users", ...)) before the English-name prompt is ever reached — not the documented query_ip_number() limitation, but a different WASM-harness gap: wasm_client.js's copyDir() deliberately only creates a bare log/ directory in the in-memory FS (skips copying subdirectory contents to avoid runtime-churn bloat), so log/login/ doesn't exist inside the WASM instance even though it's a real, populated directory in the native work/ tree. The unguarded write_file() in logind.lpc (not the hardened log_file() simul_efun) throws, and the connection is dropped. Not a mudlib bug — natively this directory exists and the counter update works fine — and not patched here (out of this pass's scope to modify the shared test harness or add defensive wrapping for a directory that's only missing under this specific test tool). Playability under WASM for this lib should be considered "boots, but blocked at the very first prompt by a harness FS gap," distinct from the true IP-gating limitation seen on other libs.

WASM-enablement pass (2026-07-23)

Standard four-change pass (AGENTS.md §1.3b/§1.3e/§1.5). Gates patched:

1. Loopback always allowed — adm/daemons/logind.lpc logon() (around line 78-95): added an _local predicate computed from query_ip_number(ob) (true for 127.0.0.1, a leading 127. prefix, an empty/non-string IP, or any IP that fails a 4-part sscanf — i.e. WASM garbage). Guarded both connection-rejection gates with it: - the BAN_D->is_banned(query_ip_name(ob)) destruct gate (line ~89), - the per-IP multi-login cap if (IPCount > 10) destruct(ob) (line ~113). adm/daemons/band.lpc's is_banned() takes a *site name* string (not a raw connection) and is only reached via the guarded logind gate, so no separate patch there. 2. Uptime startup gate — none in the login path (no uptime()<N connection gate exists in logind); nothing to bypass. 3. Anti-flood throttle — the only per-IP throttle is the IPCount>10 multi-login cap, now loopback-exempt via the same _local guard. 4. Admin account seeded — id fluffos, pw Mud@2026, name 浮浮, registered through the normal flow; granted (admin) by adding fluffos (admin) to adm/etc/wizlist. Verified update works. Save file: work/data/user/f/fluffos/{user.o,login.o} (NOT gitignored; shows as untracked — orchestrator must git add it).

Retest: fresh registration (秦风, race human, male) reaches world, score renders; fluffos login + update works; zero new runtime errors in debug.log. Test char freshtest removed; fluffos kept.

Fail-closed loopback retrofit (2026-07-24)

Security correction, applied retroactively. The original loopback predicate above (item 1) treated an empty/non-string/malformed IP as "local" (fail-open) — a defensive stopgap for a since-fixed WASM driver bug (query_ip_number() used to return garbage like "(" on WASM connections). That bug is now fixed upstream, so this lib's _local check was tightened to fail-closed: adm/daemons/logind.lpc's logon(), _local is now stringp(_ip) && (_ip == "127.0.0.1" || _ip == "::1" || (strlen(_ip) >= 4 && _ip[0..3] == "127.")) — an unparseable/empty IP is now treated as remote/untrusted (subject to the normal ban/multi-login gates), not as loopback. Retested: fresh registration (id gatetesterone→gatetestone, name 秦岭客, race human, male) still reaches the world via loopback with look/score/quit all correct; fluffos login + update /adm/daemons/logind still succeeds. Zero new runtime errors from this change (the pre-existing apply() with insufficient permission ... command_hook ... needs: private, has: hidden lines seen in debug.log are the known §8.3a private nomask command_hook class in feature/command.lpc — NOT introduced by this pass, and did not block look/score/update/quit in any observed session here; flagging for a future pass rather than fixing now since it's outside this pass's four-item scope).

深度功能测试（第二轮，2026-08-03）

之前的会话只做过注册流程 + 管理员权限验证，本轮做了完整的 §10.7 深度功能测试。proactive 检查 AGENTS.md 已归档的四类常见坏味道：

• §8.3a private nomask command_hook：只在死档案 feature/command.c.new/feature/command.bak 里还留着，真正生效 的 feature/command.lpc 已经是 nomask int command_hook（没有 private）——查了 git log 确认这是本项目更早一次系统性排查 （20f407c4048）就已经修好的，上面"已知但未修复"那条关于 apply() with insufficient permission ... command_hook 的记录已 经过时，本轮重新测试全程没有再出现这条报错——更正：这条不再 是待办事项。
• 命中了新的一类：logind.lpc 里两条并行取名流程（接受系统随机中 文名 / 自己手动输入）都各自残留一行调试用的 printf("%O\n", ob)， 紧挨在"请设定您的密码："提示之前，每一个新玩家注册时都会看到 （AGENTS.md §7.34，新增 xkyx3b 到确认实例列表）。已删除两处。
• 未命中 unguarded MESSAGE_D->、stat/water 键名、§8.9 坏 age 判断 这三类。

完整流程验证：连续注册多个全新角色，走完整个官方新手教程（游 戏本身把教程做成了六个连续房间：start1如何移动→start2如何看/ 捡→start3如何吃喝购买→start4如何求助→start5关于战斗→down 进入正式世界），教程房间之间的出口方向并不是简单的连续东进（ start2出口是south，start3出口是west，start4出口是north）， 按各房间实际定义的方向走通；start5教程明确指引"wear cloth"/ "wield sword"装备防具武器后"fight dog"和野狗打一场——照做，完整 交手十余回合，伤害叙述随野狗伤势分级变化（从"可能受了点轻伤"一路 到"受了相当重的伤，只怕会有生命危险"再到力竭倒地），战斗系统工作 正常；score/hp面板渲染正确（食物/饮水槽创建时即为满值，没有 §8.9 症状）；一路到达"中央广场"（登录留言板显示15张留言）， quit干净退出（"欢迎下次再来！"）。全程 debug.log 没有任何一条 执行时段错误——是这次会话覆盖的几个 lib 里少见的、从注册到教程 到战斗到 quit 全程零报错的一次。

未覆盖范围：留言板内容、拜师、商店购买、种族差异（矮人/精灵/ 妖精/龙人/兽人五个非人类种族本轮未测）因时间原因未实测。

Round-four gap-closing pass: board/apprentice/shop/race (2026-08-24)

Closed all four gaps flagged above. scripts/tmux_mud.sh hit its known telnet-escape-character artifact partway through the board test (a Chinese UTF-8 byte sequence tripped the *local* telnet client's ^] command mode, not a mudlib bug) — switched to scripts/mudclient.py (raw socket, no local telnet interpretation) for the rest of the session, which had no such issue.

• Message board: post <title> → in-line editor (end with .) → look/read <n> all worked correctly on /d/center/guangchang's board (count went 15→16, title/author/body all round-tripped correctly); cleaned up with discard 16 afterward. Clean.
• Shop (buy): /d/beginner/start3's waiter NPC (F_VENDOR mixin) — granted admin 10 两白银 (1000 value) via clone + call ->set_amount(), then buy long sword (id price 5 两白银/500): correct item added to inventory, correct change left (500, i.e. 1000-500). MONEY_D's change-making arithmetic in adm/daemons/moneyd.lpc is correct. Clean.
• Sect apprenticeship (拜师): real master NPCs (master-zhang 武当派, hong-qigong 丐帮, etc.) all gate attempt_apprentice() on skill levels far beyond what a fresh character has (50-150+), so admin used call me->set(...)/set_skill(...) to satisfy master-zhang's requirement (cps>=20, con>=25, int>=34, taiji-sword>=50, taiji-force>=60), then apprentice zhang from /d/wudangshan/wd-zl-4: full flow worked end-to-end (NPC accepts, auto-recruits, kowtow messages, "恭喜您成为武当派的第二代弟子"). Reverted the admin's stats/skills/family/title back to pre-test values afterward. Clean — no crash anywhere in the apprentice/recruit code path.
• Non-human race creation: fresh registration (id xkyxraceb, name 秦晓月, race 2=精灵/Elf) all the way through look/score/quit. 种族：精灵族 displayed correctly, title 精灵族普通百姓, and all eight base stats matched the elf stat block in logind.lpc's init_new_player() (13,14,21,21,14,20,15,17-shaped, ± the 0-2 random per-stat bonus set_attribute() adds) — no crash, no mismapped stat. Test character's save deleted afterward.

Bug found and fixed (real crash, confirmed live in debug.log, not content): wandering NPCs (巡捕/捕快, xun-bu) periodically drift into rooms that haven't been visited yet, triggering that room's first reset() → make_inventory() for its "objects" mapping. A handful of rooms/NPCs reference files that don't exist in this archive:

• /u/bibi/hai, /u/bibi/baotu, /u/bibi/shounao, /u/bibi/xiaoruer — items/NPCs from a wizard's /u/bibi/ home directory not included in this archive (referenced from d/center/furen.lpc, d/center/zhu.lpc, d/quanzhou/main-e2n3.lpc's "objects" mappings).
• /obj/weapon/ling (铜铃) — compiles, but its own inherit LING (/std/weapon/ling) doesn't exist anywhere under std/weapon/ (only sword/blade/axe/dagger/fork/hammer/multi/staff/stick/ throwing/whip base classes are present) — referenced from d/tianshigu/ts-dyl.lpc.
• /u/bibi/yitian (倚天剑) — same missing-home-dir shape, but reached via an unguarded chained call instead of the "objects" mapping: d/wizard/npc/{shizhe,shenxian-shizhe}.lpc's create() does carry_object("/u/bibi/yitian")->wield(); — carry_object() already guards its own new() and returns 0 for a missing file, but the immediate ->wield() chain on that 0 crashes anyway.

Root cause in the shared code: std/room.lpc's make_inventory() did ob = new(file); ob->move(this_object()); with no check that new() succeeded — for an ordinary missing file this is Bad argument 1 to EFUN call_other() ... Got: int(0), and for a file that exists but has a bad inherit (the ling case) new() itself throws (Inherited file '.../ling' does not exist!) rather than quietly returning 0. Fixed by wrapping the new() in catch() and returning 0 on failure; reset()'s case 1 branch (single-instance "objects" entries) also lacked the same "skip if the object didn't load" guard that the default: (multi-instance) branch already had via its continue, so added a matching if (!objectp(...)) break; there too. Also fixed the two carry_object(...)->wield() unguarded chains in shizhe.lpc/shenxian-shizhe.lpc directly (assign-then-check instead of chaining).

Also fixed a hard compile error discovered via the same lpcc_check.sh pass, unrelated to the missing-/u pattern: d/beijing2/zhang/npc/wife3.lpc's create() called return_home("/beijing2/zhang/room7.lpc") — a bare string — but return_home(object home) (this file even carries its own byte-for-byte copy of the parent std/char/npc.lpc implementation, doing nothing custom) requires an object. This is dead/erroneous leftover code (the NPC already manages its own coming-and-going via goout()/do_return() with hardcoded move() paths); the call has no effect on anything else in the file and its bad-type argument was a hard compile failure that took the whole NPC object (and by extension its room's "objects" entry) down. Removed the line.

Verification: scripts/lpcc_check.sh pass count went from 3116→3142/3178 (fail 62→36) purely from these fixes (no other changes) — the room-load crash was cascading into several other objects' compile counts. Live-boot spot check: fresh driver restart, visited /d/center/furen, /d/center/zhu, and /d/wizard/guest_room (where shenxian-shizhe really lives) directly — all load cleanly, shenxian-shizhe's missing sword silently and correctly skipped (just wears the cloth, no crash). Zero new 执行时段错误 in debug.log from normal play afterward. (One artifact-only error was seen mid-session from goto <bare npc file> directly instantiating shizhe.lpc outside any room, so its greeting() → command("say ...") hit a null environment() — not reachable in normal play since shizhe.lpc isn't placed in any room's "objects" mapping at all, and shenxian-shizhe.lpc (which is) works fine when visited the normal way.) Remaining lpcc failures are pre-existing, individually-untriaged content gaps (missing globals, syntax typos, plus the still-irreducible /obj/weapon/ling compile failure since its /std/weapon/ling base class was simply never included in this archive and authoring a new weapon-type class is out of this pass's scope).

WASM 修复摘要（迁移自 meta.json 的 group_note）

状态已从过时的 limited 修正——这份档案自己的 README 和 group_note 里从未记录过任何缺陷说明，本轮重新测试也没有发现：管理员登录（fluffos/Mud@2026）干净正常，'目前权限：(admin)'，quit 正常。

§7.86 跨库扫描修复（留言板 post 崩溃）

• BULLETIN_BOARD inherit + 多余 replace_program() 致命形状（AGENTS.md §7.86，post 命令崩溃）：全档案 4 处命中，已删除多余的 replace_program(...) 调用（保留 inherit），逐文件保留原有行尾格式（CRLF/LF 按文件原样）。本次为跨库 §7.86 扫描修复（触发原因：该 bug 已在 6+ 个互不相关的血统家族独立确认，属于近乎普遍的拷贝粘贴模式），仅做编译检查（驱动干净启动、端口正常监听），未做完整 §10.7 深度游玩测试。

信件系统整体死亡（与 xkyxciii 完全同源的 bug）

adm/daemons/logind.lpc 的 enter_world() 在创建/挂载玩家的邮箱物件 之后紧跟一行多余的 destruct(mail_box)——这销毁了那个在 init() 里注册全部六个信件指令（send/mail/forward/check/from/ read/discard）的物件，导致每一次登录、每一位玩家的信件系统 完全失效，报"什么？"（未知指令），且没有任何玩家可见的错误提示。 这个 bug 是 xkyxciii 那一轮 round-three 深度测试发现并修复后， 用同一行代码做全库 grep 时命中的姐妹档案（libs/xkyxciii 的 NOTES.md round-three 章节里已经记录了这条线索）。修复：删除多余的 destruct(mail_box) 调用。已用干净重启的驱动实测验证： fluffos/Mud@2026 登录后，send/check/from/read 全部给出 正常的交互式提示（标题/内容/是否留底 y-n 询问、"没有这个编号的 信件"等），不再是未知指令；debug.log 全程干净。

§7.100 sweep (2026-08-19): redundant replace_program(ROOM); landmine

Same corpus-wide bug as documented at AGENTS.md §7.100: rooms inheriting ROOM (/std/room) had a redundant, harmful replace_program(ROOM); call right after inherit ROOM; in create(), setting a permanent "pending replace" flag that crashes the object the first time anything binds a closure to it. This lib had 1,452 live occurrences (survey-ranked #90 of 166 candidates >=100, tied with sibling lib xkyxciii, both 侠客英雄传III but distinct archives/ snapshots). Fixed with the sweep's binary-mode script (fix_710_room.py); git diff --numstat totals (0 insertions, 1452 deletions) match the survey's live-occurrence count exactly. Like xkyxciii, this lib has no in-game room-building tool, so no factory-bug variant to fix. No work/data/ room-source false-negative found. Verified via a clean build-debug boot (zero "cannot replace"/"cannot bind" debug.log lines, port 40036 listening) plus a live spot-check: no admin fluffos save currently exists in data/login/, so a fresh test account was registered through the full flow (id → confirm → Chinese name → password → email → gender → race) and landed at the new-player training room; look/quit both worked, debug.log stayed clean throughout. Test account's save directory deleted before committing.

§7.30 uninitialized-mapping accessor sweep (2026-08-20)

Corpus-wide mechanical sweep of the feature/skill.lpc shared-lineage bug (confirmed independently on xiakexing2017/jqxz2015/haiyang2 via round-four testing): 3 accessor(s) in this file returned a raw never-initialized mapping instance variable (defaults to int 0, not ([]), until first assigned), crashing any unguarded keys()/sizeof()/indexing caller for a fresh/untrained character. Fixed at the accessor level (mapp(x) ? x : ([])) per the documented remedy. Verified via lpcc --batch static compile check only (not a live boot) as part of a large mechanical sweep; not individually functionally re-tested live on this lib.