ROUND2.md (10690B)
1 <!-- 2 This file is part of GNU Taler 3 (C) 2026 Taler Systems S.A. 4 5 GNU Taler is free software; you can redistribute it and/or modify it under the 6 terms of the GNU General Public License as published by the Free Software 7 Foundation; either version 3, or (at your option) any later version. 8 9 GNU Taler is distributed in the hope that it will be useful, but WITHOUT ANY 10 WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR 11 A PARTICULAR PURPOSE. See the GNU General Public License for more details. 12 13 You should have received a copy of the GNU General Public License along with 14 GNU Taler; see the file COPYING. If not, see <http://www.gnu.org/licenses/> 15 --> 16 17 # POS simulator, round two: fault injection and the other layout 18 19 Round two set out to raise coverage, with `taler-twister` for the network side. 20 It ran in two passes. The first used five emulators, which `tripwire` cannot 21 carry -- load reached 10 and the time per scenario went from about ninety 22 seconds to several minutes -- and was abandoned half finished. The second used 23 three, which is what the box takes: load settled around three and the pace came 24 back. Everything below is from the second pass plus what the first had already 25 collected. 26 27 ## Coverage 28 29 ``` 30 round one round two 31 instruction 31097/40234 77.3% 34219/40234 85.0% 32 branch 1063/1739 61.1% 1159/1739 66.6% 33 line 3827/4864 78.7% 4155/4864 85.4% 34 complexity 1111/1897 58.6% 1217/1897 64.2% 35 method 792/1047 75.6% 835/1047 79.8% 36 class 123/140 87.9% 130/140 92.9% 37 ``` 38 39 `+328` lines and `+96` branches, and **no class is at 0% any more**: round 40 one's one untouched class, `HistoryManager$forceDeleteOrder$1`, is reached 41 now. The round-one dumps are merged in: the APK is 42 byte-identical, and jacoco matches execution data to classes by checksum. 43 44 Where it went: 45 46 ``` 47 ProcessPaymentFragmentKt 42.5% -> 93.4% (+115 lines) 48 AmountEntryFragmentKt 78.7% -> 95.6% (+46) 49 OrderFragmentKt 76.4% -> 84.7% (+51) 50 PaymentManager 62.1% -> 83.5% (+22) 51 ConfigManager 71.3% -> 82.2% (+28) 52 ~.payment (package) 62.5% -> 90.5% 53 ``` 54 55 Branch coverage rose by five and a half points against line coverage's six and 56 a half, where round one had branch coverage lagging line coverage by 57 eighteen. That is what the 58 fault scenarios bought: the lines were already being executed, and what was 59 missing was the arm taken when the backend misbehaves. 60 61 ## What moved, and why 62 63 ### A phone-sized emulator 64 65 Five of the terminal's screens choose their layout at runtime on 66 `smallestScreenWidthDp >= 720` -- the order grid, the payment screen, the 67 refund URI screen, the QR config tab and the amount keypad (the last on 68 `>= 600`, with a third arm for a short screen). Round one ran a tablet only, 69 so the phone half of each was never executed, and in three of those files it is 70 the larger half. 71 72 The device profile does not decide this: `emu.sh` writes the LCD geometry into 73 the AVD and that is what the app sees, so a `medium_phone` AVD still came up 74 2560x1600 at 320dpi -- 1280x800dp, a tablet. `POS_LCD_*` settles it, and 75 1920x1080 at 420dpi is 731x411dp. 76 77 ### Twister: a real backend made to go wrong 78 79 `driver/faults.py` puts `taler-twister` -- the fault-injecting proxy the C test 80 suites use -- in front of `backend.demo.taler.net`, and `driver/suite.py` gains 81 thirteen `fault` scenarios. Everything stays real except the one field, status 82 code or byte count that was armed. Two things had to be bridged: 83 84 * **TLS.** Twister is a plain HTTP proxy and the terminal hard-codes `https://`. 85 `faults.TlsFront` terminates TLS with the certificate the mock already uses 86 and relays bytes -- a byte relay, not an HTTP proxy, so keep-alive, chunked 87 bodies and uploads pass through untouched. 88 * **One mutation, one request.** A twister mutation is consumed when it fires, 89 which suits a test script and not a screen that polls: the insecure-pay-URI 90 dialog opens on one poll of the order status and the next poll closes it. 91 `Twister.sustain()` re-arms in a loop. 92 93 `TlsFront` also injects what twister has no notion of because it is below HTTP: 94 a body that stops halfway, or a reset while the request is still going up. 95 96 ### A mock backend that demands a two-factor code 97 98 Round one reported the two-factor login path as unreachable: the demo backend 99 has `mandatory_tan_channels: ["email"]` and mails the code. It is reachable 100 after all -- a mock can be that backend with a code the harness knows, and the 101 terminal cannot tell the difference. The whole dance is four requests and one 102 header, and `ChallengeRequiredException`, 103 `ConfigFragment.fetchLimitedAccessTokenWithMfa` and `ConfigManager`'s challenge 104 calls are now covered. 105 106 Three scenarios walk it: one channel chosen from two; three wrong codes and 107 then the right one, which covers both statuses the terminal treats as 108 recoverable (409 "wrong code" and 429 "a new code is on its way"); and 109 `combi_and`, where both channels must be answered and no chooser is shown. 110 111 ## Findings 112 113 | | Finding | Severity | 114 | --- | --- | --- | 115 | 10 | [A connection that dies mid-response is reported in the HTTP engine's words](transport-error-shows-the-library-wording.md) | low-medium | 116 | 11 | [A missing field is reported to the cashier by its Kotlin type name](serialization-error-shown-to-the-cashier.md) | low-medium | 117 118 Both are the same arm of `Response.getFailureString` -- the one that shows 119 `exception.message` -- and both are reachable only through a proxy over a real 120 backend. A cut connection produces OkHttp's "unexpected end of stream"; a 121 deleted field produces kotlinx.serialization's "Field 'orders' is required for 122 type with serial name 'net.taler.merchantlib.OrderHistory'". Round one's nine 123 stand. 124 125 Also observed, and not promoted to a finding because it has not been checked on 126 a real device: 127 128 * **The login form does not fit a phone in landscape.** At 731x411dp the form 129 ends below the bottom edge: Password is clipped, and Connect and the 130 "No instance?" help card are off-screen, with nothing to say the screen 131 scrolls. The terminal is landscape-locked, so turning the device is not 132 available. The driver now scrolls; a person would have to work it out. 133 134 ## What the terminal got right 135 136 * **A sold-out product.** A 410 carrying `product_id` and the quantities is the 137 one backend error the terminal translates rather than quotes, and it does: 138 "Error: Not enough inventory available / Inventory stock unavailable for 139 product ...". It reloads the inventory afterwards, as it should. 140 * **The insecure payment link**, from all three sides. Twister handing out a 141 `taler+http://` URI raises "Payment link problem", and "Try fixed link" 142 produces a `taler://` URI, "Continue as is" keeps the insecure one, and 143 "Cancel order" leaves the payment screen. The wording names no scheme, which 144 is the right call for a cashier. 145 * **Two-factor login**, including three wrong codes in a row: the terminal 146 asked again each time, requested a fresh code when the backend rate-limited 147 it, and sent the code hyphenated as `0000-0000` the way the backend expects. 148 * **An order whose reply was truncated.** The order really was created; the 149 terminal reported the failure and left no second order behind. 150 151 ## False positives, found and removed 152 153 Four this round, each confirmed by re-running the scenario against the 154 corrected harness: 155 156 | Reported | What it really was | 157 | --- | --- | 158 | `zero-amount-accepted-silently` | at 0.00 the charge button is *disabled*, which is better than a dialog; the oracle insisted on the dialog | 159 | `keypad-overflow-blank` | the keypad groups thousands, so its largest amount reads `999,999,999,999.99` and the parser only knew the ungrouped shape | 160 | `insecure-uri-not-honoured` | the URI was read after `sustain` stopped re-arming, so the backend was healthy again and the terminal had correctly re-read it | 161 | `insecure-uri-dialog-says-nothing` | the dialog says "payment link" rather than "http", deliberately; the oracle was looking for the jargon | 162 163 Four harness bugs behind them, all fixed: 164 165 * **Back dismisses a Compose dialog**, and `tap_button` used Back to put the 166 keyboard away before every tap -- so the custom-product `Add` and the 167 two-factor `OK` disappeared before they could be pressed. The keyboard is now 168 only put away when the dump's own root bounds say it covers the control. 169 * **One wallet shared by three workers.** They spent each other's coins; the 170 symptom was one worker topping up to 60 and another failing with 171 "insufficient balance" in whatever scenario came next. One wallet and one 172 bank account per worker now. 173 * **A dialog covering the next tap.** A scenario that arms a fault, reads the 174 dialog and arms the next one has to dismiss it first, or it waits out its 175 timeout looking for a button behind a modal. 176 * **Scrolling a control out of existence.** `_scroll_into_view` could move a 177 dialog's own list and carry the target off screen; losing it is worse than 178 not having moved it, so that is undone. 179 180 ## What round three should do 181 182 1. **The refund form on a phone**, which is the one open question this round 183 leaves. `refund_full` and `refund_amount_edges` both fail there, twice each, 184 and the screen they fail on carries only its title and the order summary -- 185 no amount field, no buttons -- while the same scenarios pass on a tablet. 186 Either the form clips its controls on a short screen, which is a finding, or 187 the driver is still mishandling the keyboard, which is not. A list of labels 188 cannot tell those apart, so `Pos.start_refund` now photographs the screen 189 when it happens and the next run settles it. `RefundUriFragmentKt` at 45.3% 190 is the largest single remainder and sits behind the same door. 191 2. **The config package**, the largest remaining gap at 81.1%: the QR 192 configuration tab is most of it, and it crashes (round one, finding 1). 193 3. ~~**`fault_login_connection_cut` is flaky through the proxy.**~~ Settled in 194 round three: it was not the proxy. A change to the driver -- putting the 195 keyboard away only when it is over the control -- stopped it being put away 196 on the login form, which does not resize under the IME, so the tap on 197 Connect landed on a key. With the keyboard handled from the window 198 manager's own inset the scenario passes and finding 10 reproduces. 199 4. **More chaos.** `fault_chaos_walk` runs the terminal under a 12% failure 200 rate and it came through clean; one walk is not many, and it is the only 201 scenario where a fault arrives at a moment nobody chose.