julien/typewriter
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
e53be5062d5dbc0742eb20c0871701b481631741
typewriter/docs/qfd.md

32 KiB
Raw Blame History

Quality Function Deployment

Translates what the device must be (user-facing requirements) into what it must do (engineering functions) and what we must build (components). Surfaces the few targets that dominate the design and the conflicts between them. Every decision cell points back to adr.md.

Scope: v0.1 MVP — see v0.1-mvp-product.md for user-facing scope and v0.1-mvp-technical.md for implementation — with the v0.2–v1.0 trajectory (README, roadmap) in mind so we don't paint into a corner. Terminology (e.g. Tracked, Local, Save, Publish) follows the project glossary at ../CONTEXT.md.

Format inspired by the classic House of Quality, kept compact. Strength weights: 9 strong, 3 medium, 1 weak, blank none.

1. Customer requirements (the WHATs)

What a user (= me) values about the device, with importance weights on a 1–10 scale. Source columns point at the doc the requirement comes from.

ID Requirement Weight Source
W1 Sub-second visible response to typing 10 product → Write, README → UX
W2 Publishing is one deliberate action away 9 product → Publish, CONTEXT → Publish
W3 Pulling power never corrupts the file 10 product → Recover, acceptance
W4 Provisioning never interrupts a writing session 7 product → Provisioning, roadmap → v0.9
W5 Quick boot to a writing cursor 6 product → acceptance (≤ 5 s)
W6 Long sessions without crash / lag / drift 9 product → acceptance (1 h soak)
W7 Nothing on the device competes with prose 8 README → vision
W8 The UI never moves except when I move it 7 README → UX
W9 Codebase absorbs the planned roadmap without rewrite 8 roadmap
W10 I can repair or fork it with hobbyist tools 5 README → vision
W11 Multi-day battery life (v0.8 onward) 4 roadmap → v0.8
W12 Local-only file scope coexists with git scope (v0.5+) 5 README → scopes, roadmap → v0.5
W13 Typography sets a writing-tool tone — typewriter or developer editor, never gadget 7 roadmap → v1.0, README → UX

2. Engineering functions (the HOWs)

Measurable characteristics. Targets are v0.1 unless noted. Direction column shows what "better" looks like (↑ higher, ↓ lower, → fixed).

ID Function Dir v0.1 target v1.0 target
H1 Keypress → glyph latency ↓ ≤ 200 ms ≤ 150 ms
H2 Partial-refresh region area per keystroke ↓ ≤ 1 text line (~22 px h) same
H3 Full-refresh cadence (clears ghosting) → 1 per 20 partials tuned by panel temp
H4 Cold boot → cursor ready ↓ ≤ 5 s ≤ 3 s
H5 Continuous-typing endurance (no drop, no leak) ↑ ≥ 1 h ≥ 8 h
H6 Ctrl-G push success rate on healthy Wi-Fi ↑ ≥ 95 % ≥ 99 %
H7 Push end-to-end (one-file commit) ↓ ≤ 30 s ≤ 10 s
H8 Save survives power loss after status confirms → 100 % 100 %
H9 PSRAM heap headroom during push ↑ ≥ 1 MB free at peak same
H10 Firmware binary size ↓ ≤ 2 MB ≤ 1.5 MB
H11 Stack budget across all tasks ↓ ≤ 80 KB (sum) same
H12 Wi-Fi reconnect on transient outage ↓ ≤ 30 s ≤ 10 s
H13 Idle / typing / push current draw ↓ measured only sized for >2 days
H14 Module count / public-API surface (refactor proxy) → ≤ 8 modules same
H15 Build time (clean, release) ↓ ≤ 7 min ≤ 5 min

3. House of Quality — WHATs × HOWs

Reading: row × column cell is how strongly the function (H) advances the requirement (W). Importance at the bottom is Σ(weight × strength) — the weighted vote on which functions deserve the most engineering attention.

H1 lat H2 area H3 cad H4 boot H5 soak H6 push% H7 push s H8 dura H9 heap H10 bin H11 stk H12 wifi H13 mA H14 mod H15 build
W1 (10) 9 9 3 3 1 1
W2 (9) 9 3 9 9
W3 (10) 9
W4 (7) 3 3 1
W5 (6) 9 3
W6 (9) 3 3 9 3 3 9 3 3
W7 (8) 3 3 3 3 1
W8 (7) 1 9 9
W9 (8) 1 1 9 3
W10 (5) 3 1 3 1
W11 (4) 9
W12 (5) 1 3 3
W13 (7) 3
Σ 148 177 144 54 111 134 27 132 193 41 45 129 65 117 29

Top engineering priorities (from importance)

  1. H9 — PSRAM heap during push (193). gitoxide pack + rope + TLS all share the same arena; ADR-001 and ADR-004 trade binary size for ecosystem so this becomes the watched metric. The umbrella typography WHAT (W13) keeps a fixed-size glyph-cache load on top of that arena pressure.
  2. H2 — partial-refresh region area (177). Bound how many pixels the panel has to flip per keypress; ADR-003 is the hardware-side answer.
  3. H1 — keypress latency (148). The single most user-visible number; ADR-002 and ADR-003 are co-conspirators.
  4. H3 — full-refresh cadence (144). The ghosting/flash tradeoff; lives in the render layer.
  5. H6 — push success rate (134). ADR-004 (gitoxide) and ADR-005 (PAT over HTTPS) own this jointly; spike 7 is the kill-switch.
  6. H8 — save durability (132). Atomic-rename + fsync; FAT's weakness is acknowledged in ADR-007 and mitigated, not designed around.

Why H8 ranks where it does. HoQ totals reward functions that touch many WHATs over functions that absolutely matter for one WHAT. W3 ("Pulling power never corrupts the file", weight 10) is the strongest single requirement but its only strong link is H8. H9, by contrast, collects from W1+W2+W6 to reach 193. The fix isn't to re-weight cells to chase intuition — it's to use §6 as a curated rank that lifts narrow-but-critical functions explicitly.

The bottom three (H7 push time, H15 build time, H10 binary size) are real costs but ones we knowingly took on (ADR-001) and are not in the critical path of user experience. The tightened H15 v0.1 target (≤ 7 min) reflects user preference for faster iteration, not matrix-derived priority; if it pushes back against ADR-001's "+5–10 min" pricing, the target moves before the runtime decision does.

4. Roof — function-vs-function tradeoffs

The roof shows where pushing one function pushes another the wrong way. Classical QFD single-character symbols: ◎ strong reinforcement, ○ mild reinforcement, × mild conflict, ⊗ strong conflict.

H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 H13 H14 H15
H1 — ◎ × ○ ×
H2 — ◎ ○
H3 — ○
H4 — ×
H5 — ○ ○ ⊗
H6 — ○ ⊗ ◎
H7 — × ◎
H8 —
H9 — ⊗
H10 — ⊗
H11 — ×
H12 —
H13 —
H14 — ×
H15 —

Conflicts that actually shape the design

  • H1 latency ↔ H3 refresh cadence (mild). More partial refreshes per second pile up ghosting faster, demanding earlier full refreshes — visible flashes that hurt H8 perception and H1 burst behaviour. The ADR-003 strip aspect is the structural answer: a small framebuffer makes both cheaper, not one at the expense of the other. The runtime answer is render §H3: schedule full refreshes on idle ≥ 1 s (v0.1 tech doc).
  • H9 heap ↔ H10 binary size (strong). std + gitoxide + mbedtls inflate both. We chose to spend on these (ADR-001, ADR-004) because 16 MB flash and 8 MB PSRAM make them affordable; the kill-switch is spike 7. If heap during push refuses to come under 1 MB free, ADR-004 flips to libgit2-sys for v0.1.
  • H9 heap ↔ H5 soak (strong). A long writing session grows the rope and the glyph cache; pushing on top can OOM. Mitigation: 256 KB file cap (v0.1 tech doc) + glyph cache eviction before push + watching the spike in spike 7.
  • H6 push success ↔ H12 Wi-Fi reconnect (reinforcing). Both come from the same network stack; investing in reconnect backoff helps both.
  • H10 binary ↔ H15 build time (strong). std builds are slow. Accepted in ADR-001 — refactor leverage (H14) is the long-term payoff, not the per-build seconds.
  • H4 boot ↔ H10 binary (mild). Larger binary = slower flash load. Affordable at our size class but worth watching as features land.
  • H11 stacks ↔ H13 current draw (mild, future). Idle threads draw little but never zero; a future light-sleep policy (v0.8) wants them parked.
  • H14 modularity ↔ H15 build time (mild). More small crates = more link work. Boring vs valuable; we lean toward modularity.
  • W13 typography ↔ H9 heap + H10 binary (mild, future). Achieving a writing-tool tone needs room for glyph caches and font assets. Not load-bearing in v0.1 (one mono font), but the v1.0 tone goal is why H9 and H10 keep slack rather than being squeezed to the minimum.
  • Tightened H15 ↔ ADR-001 (mild). Pulling v0.1 build time from ≤ 10 min to ≤ 7 min eats into ADR-001's accepted "+5–10 min" cost. Worth aiming at via cargo profile / vendor LTO / crate-graph trims; worth giving up before reversing ADR-001.

5. Function → Component mapping (Phase 2)

Which subsystem owns the delivery of each function. Cells are which ADR constrains the choice.

Components (with anchoring ADR):

ID Component ADR
C1 ESP32-S3-N16R8 SoC ADR-001, ADR-008
C2 esp-idf-rs (std) + ESP-IDF ADR-001
C3 std::thread + crossbeam-channel ADR-006
C4 PSRAM allocator wrapper ADR-001
C5 GDEY0579T93 + DESPI-c579 panel ADR-003
C6 embedded-graphics + e-paper driver ADR-002, ADR-003
C7 Custom widget / dirty-rect layer ADR-002
C8 ropey rope buffer ADR-001 (ecosystem)
C9 TinyUSB host (esp-idf bindings) ADR-009
C10 FAT on microSD ADR-007
C11 LittleFS on internal flash ADR-007
C12 gitoxide (gix-*) ADR-004
C13 mbedtls TLS (via ESP-IDF) ADR-005
C14 HTTPS + GitHub PAT auth ADR-005
C15 eFuse-derived encryption key ADR-005, ADR-007
C16 USB-C wall PSU ADR-008

Function-to-component matrix (9 strong / 3 medium / 1 weak):

C1 SoC C2 std C3 thr C4 PSR C5 EPD C6 eg C7 wid C8 rope C9 USB C10 SD C11 LFS C12 gix C13 TLS C14 PAT C15 efs C16 PSU
H1 lat 3 1 9 3 9 9 9 3 9
H2 area 9 9 9
H3 cad 9 3 9
H4 boot 3 9 3 1 3 9 3
H5 soak 3 3 3 9 1 9 9 3 3 3
H6 push% 3 9 9 9
H7 push s 3 1 3 9 9
H8 dura 3 9 9
H9 heap 3 3 9 3 9 9
H10 bin 9 1 3 3 3 3 9 3
H11 stk 9 3 3
H12 wifi 3 9 3
H13 mA 9 1 9 3 3 9
H14 mod 3 3 3 9 3 9
H15 build 9 9 3

Read across, not down

  • C5/C6/C7 (panel + graphics + widget) are the single most leveraged cluster — they own H1, H2, H3 (the top of the priority list). ADR-002 and ADR-003 are the ADRs to keep most honest as v0.x progresses.
  • C12 (gitoxide) is overloaded: H6, H7, H9, H10, H11, H14, H15 all touch it. That's why ADR-004 includes a kill-switch (fall back to libgit2-sys if spike 7 fails). It's also why H9 sits in the top three priorities — gitoxide's memory profile is the unknown. ADR-010 pins the shape of the publish sequence (the gct flow); C12 is just the library that implements it. Changing ADR-010 doesn't change C12's column, but changing C12 (the kill-switch) does not change ADR-010's user contract.
  • C11 (LittleFS) is unused in v0.1 — config is build-time. Its non-zero cells in the matrix describe the v0.9+ shape per ADR-007, not v0.1 reality.
  • C2 (std runtime) sits underneath almost everything, but it's the enabler (H4 boot, H10 binary, H12 Wi-Fi) rather than the bottleneck. Reversing ADR-001 would force re-deciding ADR-004, ADR-005, ADR-006, ADR-007 all at once — they're a single decision in three drawers.

6. Critical performance budget

A curated rank, drawing from §3 importance and §4 conflicts but with two deliberate overrides: (a) acceptance-criteria critical paths (H4 boot, H5 soak) move up regardless of weighted-vote spread, and (b) table-stakes correctness (H8 durability) moves up despite a narrow voter base. These are the numbers spikes 2–7 must validate before integration starts.

Rank Function Target Watched on If we miss it
1 H2 region area ≤ 1 line per keypress spike 2 + spike 5 Increase font size to shrink per-glyph dirty rect (ADR-003 consequence)
2 H9 PSRAM heap ≥ 1 MB free at push peak spike 7 ADR-004 kill-switch → libgit2-sys; cap rope at 128 KB
3 H8 durability 100 % survive power yank after status bench HIL Re-evaluate ADR-007 (move config to internal NVS only)
4 H1 latency ≤ 200 ms keypress→glyph spike 5 Larger partial-refresh region; render multi-char bursts
5 H6 push % ≥ 95 % on healthy Wi-Fi spike 6 + spike 7 TLS cipher trim; reconnect backoff tuning
6 H3 cadence full every ~20 partials spike 2 Adjust per panel temperature; defer flash to idle ≥ 1 s
7 H4 boot ≤ 5 s to cursor integration smoke Trim startup logging; lazy-mount SD after splash
8 H5 soak 1 h no leak / no drop 1 h bench soak Glyph-cache eviction; PSRAM heap-fragmentation review

The two not-in-MVP rows but already-shaped-by-design:

| — | H13 current | Measured only in v0.1 | bench multimeter | Cell sizing for v0.8 is data-driven, not spec-sheet | | — | H11 stacks | Sum ≤ 80 KB | static analysis | Was off-by-2x in ADR-006 pre-fix — corrected in §7 |

7. Tradeoffs and their why, linked to ADRs

Plain-language summary of what we accepted in exchange for what.

Tradeoff Got Paid ADR
std (esp-idf-rs) over no_std (esp-hal) Heap, threads, VFS, mbedtls, gitoxide-compatible +1 MB binary, +5–10 min builds ADR-001
Custom widget layer over Ratatui Dirty-rects aligned to e-ink regions; 200 KB binary back 500 LoC we own and maintain ADR-002
e-ink medium over FSTN / memory LCD / OLED Paper aesthetic; 0 W idle persistence; medium enforces writing posture ~200–300 ms typing latency; periodic full-refresh flash (scroll worst-case) ADR-003
gitoxide over libgit2-sys Pure Rust, modular, no FFI cross-compile pain Smart-HTTP path is newer; PSRAM profile unproven (spike 7) ADR-004
HTTPS + PAT over OAuth device-flow or SSH Simplest auth that gitoxide smart-HTTP already supports Long-lived secret on device; in v0.1 the PAT is compiled into the binary (dev-only target user makes this acceptable); v0.9 moves it to encrypted NVS ADR-005
std::thread over embassy or tokio Boring, debuggable, real stack traces; no exec to tune ~76 KB total stack across 5 tasks ADR-006
FAT-on-SD + LittleFS-on-flash split Desktop can read SD; config survives SD reformat Two filesystems to manage; FAT's power-loss weakness mitigated by atomic-rename ADR-007
Wall power for v0.1, battery deferred Measure real draw before sizing the cell Tethered MVP; not the final aesthetic ADR-008
USB host (TinyUSB) over BLE-HID No radio contention with Wi-Fi during push; keyboard powered from the device One more USB connector on enclosure ADR-009
Atomic Ctrl-G + auto-timestamp commit message One key, one outcome; matches the user's existing gct workflow; no modal prompt to slow H1 latency Commit history is timestamp noise; the device may author merge commits the user never sees; reversal would break muscle memory ADR-010

Conflicts left explicitly unresolved by v0.1

These are the live tensions we are watching, not deciding harder:

  • ADR-004 vs H9. If gitoxide cannot keep ≥ 1 MB PSRAM free at push peak, we are committed to switching transports for v0.1, not absorbing the OOM risk.
  • ADR-009 vs H6/H13. If TinyUSB host turns out unstable (spike 4), BLE-HID is the documented fallback — at the cost of Wi-Fi radio contention during push (re-checking H6).
  • ADR-007 vs H8. Power loss between FAT rename and dir flush yields the previous saved version. We document this as expected behavior; it becomes a real bug only if soak testing shows it triggering on routine saves.
  • W13 typography paths. v0.1 ships one mono font; v1.0's writing-tool-tone outcome admits two paths (mono = developer comfort, serif = typewriter feel). Not yet decided whether to ship both or one; decision deferred to the v1.0 design pass. Cost preview per added font: +H9 glyph-cache footprint, +H10 binary for embedded assets.

8. Inconsistencies spotted and fixed

  • ADR-006 stack figure. ADR-006 previously said "~40 KB of stack space for task stacks" — but the v0.1 technical design's task table (usb 8 + wifi 8 + ui 16 + render 12 + git 32) sums to 76 KB. Updated ADR-006's Consequences section to reflect the actual budget and cross-reference the tech doc. The 76 KB figure still fits comfortably in the ESP32-S3's 512 KB internal SRAM, so no design change — just documentation accuracy.
  • Commit-message format triple-mismatch. README said git commit -m "wip", the v0.1 product doc said "wip <timestamp>", and the user's actual shell alias (gct / git-commit-timestamp) uses a pure ISO-8601 timestamp with no wip prefix. Resolved by aligning all docs on gct and recording the decision as ADR-010. Pulled the v0.7 roadmap item "Commit message prompt instead of hard-coded wip" — it's now contradicted by ADR-010 and removed.
  • First-run flow vs. target user. The v0.1 product doc described a captive-portal first-run, but the same doc names the v0.1 target user as the dev themselves ("Me. Solo."). Provisioning a solo-dev device through a captive portal is ceremony without a user. Resolved by switching v0.1 to build-time env-var config (no NVS, no LittleFS, no AP mode); on-device provisioning is the v0.9 release that introduces non-dev users. Touches ADR-005, ADR-007, the v0.1 product + technical docs, and the v0.9 roadmap entry.
  • Vocabulary leak. Earlier docs used "commit" and "push" as if they were distinct user actions; the gct/ADR-010 model collapses them into a single user-facing Publish. Resolved by introducing CONTEXT.md as the canonical glossary; user-facing text now uses Save and Publish only.
  • House of Quality column sums recomputed. Earlier Σ row drifted from the matrix arithmetic — H1 listed 138 but sums to 148; H8 147 vs 132; H9 162 vs 172; H13 74 vs 65; smaller deltas elsewhere. Recomputed all sums from the cells. H8 dropped from #3 to #6 — a "fewer WHAT voters" artifact, not a signal that durability matters less to the design.
  • W13 reframed, W14 removed. Earlier W13/W14 rows named solutions ("beautiful monospace", "beautiful serif") inside the requirements column, conflating what the user values with which asset delivers it. Replaced with one outcome WHAT — typography sets a writing-tool tone — and moved the mono+serif option to §7 as a v1.0 unresolved tension. Σ shifted (H9 205→193, H2 198→177, H1 155→148) because the prior W13/W14 cells were scoring solution-fit rather than outcome-fit.
  • WHATs swept for solution-shape phrasing. Following the W13 reframe, the same drift was found in W2 (named the key Ctrl-G), W4 (named the process shape "one-shot"), W7 (named the hardware "surface"), W8 (named the medium "e-ink"), W10 (named the deliverable "BOM"), and W9 ("nine releases" — brittle vs roadmap reshuffles). All rephrased as outcomes; the named solutions remain documented in §7 tradeoffs and the relevant ADRs where they belong. Matrix cell strengths held — each cell scored the function against the underlying outcome, not the surface phrasing — so no Σ recompute.
  • §3 vs §6 priority lists clarified. The two were giving different orderings without saying why. §6 now states explicitly that it is a curated rank with two named overrides over §3's pure arithmetic: acceptance-criteria critical paths (H4, H5) and table-stakes correctness (H8) get manual lifts. §3 now names the HoQ structural bias that makes the curation necessary — reward for spread, penalty for narrow-but- critical functions — using H8/W3 as the canonical example.

The minor variance between README's "~12 lines" and product/ADR-003's "~11 lines" of edit area is within rounding for a 14 px glyph in a 240 px tall edit region and is not load-bearing.

How to keep this document honest

  • When a new ADR lands, add its components to §5 and re-score any function-row whose dominant component changed.
  • When a spike returns numbers, update §6's "Target" or "Watched on" columns — this is the doc that should feel out of date if measured reality drifts from estimates.
  • The WHATs change rarely; the HOWs change with each release; the matrices are recomputed when either side changes.
Reference in New Issue View Git Blame Copy Permalink