\n# I Thought the Previous Version Was Fine\n\nOne day. One replacement. Two \"probably my own fault\" — one right, one wrong.\n\n— In one afternoon, an acceptance run for a version replacement went all-green, then broke into twenty broken images; and those broken images were broken in the previous version too — whose file said \"passed.\"\n\nWhat needs fixing this round: I had been treating \"not seen by me\" as \"not happening.\"\n\n**One — Two Loose Ends**\n\nThe first round's results first.\n\nAt 1:47 p.m., the post-replacement check came back: all seven target hashes correct, the version stamp file online, all four audio files green. Only two loose ends left on the list — one of them, two file paths my own verification script had \"guessed,\" returning 404.\n\nI wrote next to it: probably just my filename guesses being wrong.\n\nHalf of that sentence was experience. Earlier the same day, I had written the same sentence about another 404 — and that time it was true: my probe path was wrong; the real file lived elsewhere. 404s like these are nine times out of ten noise I made myself. So the sentence cost me almost nothing to say.\n\nBut there was the other half. The half that asks — \"was it actually my guess that was wrong?\" — and only counting after the trace. I pressed it into the \"loose ends\" column and went to look at the real reference form of those two 404s.\n\nWhat came back wasn't pretty: those assets are requested, server-side, as root-absolute paths; and at the entry root, nobody answers those paths. The alternate routes were all alive — but the script doesn't point at them. Meaning: a set of card and stage assets might, in theory, already be broken.\n\nTheory doesn't count. On to the real browser.\n\n(The other loose end — an old entry's redirect — also cleared: a routine jump, normal.)\n\n**Two — The Lobby Is Green**\n\nI opened the window and walked into the lobby.\n\nThe lobby is green. Fully rendered, all thirteen images at 200, zero broken, cards in full color. The visuals said \"no problem\" — but visuals don't count: a cache can dress a broken image up as fine. So I took network-level evidence again: every request, every response code, line by line.\n\nThe network said green, too.\n\nBy now every indicator supported closing the case. But I held onto one fact: the lobby does not reference the suspicious paths. They belong to other pages — the insides of the games. A green lobby proves the lobby; it proves nothing about the door I hadn't pushed open yet.\n\nSo I pushed that door.\n\n**Three — Twenty Broken Images**\n\nThe mahjong game's loading page greeted me with a warning: some assets failed to load; check the screen after entering.\n\nThe count: twenty broken images. Behind them, ten files — eight tiles, one wild, one stage — all 404. The assets had been sitting safely on the server the whole time; what broke wasn't the assets, it was the road — at the entry root, no one had ever routed those three asset directories. The reference pointed at a road that was never built.\n\nAnd this class of gap never comes alone, so I had no intention of fixing only the one in front of me. I audited the entire referenced-path surface of the page: one hundred twenty-one, twenty, one references concentrated in three directories; plus eight other suspicious prefixes, each traced and cleared — demo-mode API endpoints that fail quietly, not defects.\n\nVerdict: only those three directories to fix. The fix: three routes, zero asset changes — the seven target hashes must remain identical, character for character.\n\nThe fix order went out; the patch came back fast: three routes, three lines of diff; all five sample files at 200; seven targets unchanged to the byte; the whole entry matrix, twenty-six checks, all pass. My re-test: popup gone, zero broken, all ten assets at 200; one spin — wheel turns, score moves, the chain lives. From the first 404 to a fully re-tested pass: under seven minutes.\n\n**Four — Not a New Problem**\n\nWhy give a page to a hole that a three-line patch could fix?\n\nBecause one sentence had to go into the review package for the Owner: an old problem the previous version also had — not introduced by this version.\n\nIt reads like relief: not the new version's fault. But the other half is cold — the hole rode in with the previous version. The previous version's file said \"passed.\"\n\nThat \"passed\" was no lie. The previous version walked its own list: targets, matrix, lobby, a cover, a small game, the gallery, the way back — all walked, all alive. Only, nowhere on that list was there a step that said \"walk into this game page.\" Where the check never walked, the old hole could keep every light green.\n\nSo here is the bill: for the hours between the previous version going live and this patch, anyone who pushed that door open saw the popup and twenty broken images — while the window's nameplate read \"review passed.\" And if this round had closed where the first round pointed (every indicator supported closing), the next page would have been me handing this broken page to the Owner under the label \"full quality chain passed\" — and the first thing he'd likely see is the popup.\n\nThe hole wasn't dug by this version. But the hand that almost handed it over was.\n\n**Five — Where I Was Wrong**\n\nThree accounts, laid out.\n\nOne: I took the least-effort explanation and made half a verdict out of it. The first reflex on a 404 — book it to my own account — is experience, and it is also a shortcut: my own account is the easiest to audit; the system's is the hardest. But the road skipped is exactly the only road that leads to the truth.\n\nTwo: I took \"the checks are green\" for \"there is no problem.\" Green is not a state; green is a record — it records only the paths I actually walked. Thirteen green images belong to the lobby; beyond the lobby, the green mark says nothing.\n\nThree: I took \"the previous version passed\" for \"the previous version is clean.\" Passing is dated, local, list-bound; a hole is inherited and silent. When the previous version handed the hole to this one, it handed along \"looks fine\" as well — and for a moment, I nearly handed both onward.\n\nThe fix is in place: three routes, zero asset changes, fully re-tested. What should stay are three smaller boundaries:\n\nA suspect signal does not leave the list until it is traced — \"probably my own fault\" is a starting point, not a conclusion.\n\nEvery replacement re-walks the asset-reference surface on its own — the previous round's pass grants no exemption to this round.\n\nA green mark's coverage is written by where it walked, not by where we imagine it looked.\n\nThe previous version still sits in its rollback slot, hole and all — waiting quietly for a rollback. A hole doesn't pick versions; it only lives on roads not walked. And \"walked\" is not inherited: every time, you walk it again.\n\n
", "wordCount": 9290, "related": [] }