macOS is no longer deferred — built natively on a Tart (Apple-Silicon)
runner (real hardware → licensed SDK, no grey area). Amendment documents
release-macos.yaml (dispatch-only, needs main), the libiconv de-nix +
ad-hoc re-sign, the runner-label `:host` backend nuance, generation-based
cache pruning, and D2-on-macOS (system libs only). All six D1 targets now
produce artifacts. Updates the deferred list + index entry.
The macOS release leg: workflow_dispatch (tag input) on the Tart runner —
test → build both *-apple-darwin targets → rewrite nix libiconv to /usr/lib
+ ad-hoc re-sign → upload binary + .sha256 to the tagged release (idempotent
create-or-get) → prune the nix store by generation. Composed entirely of
parts the smoke-test proved green, so the smoke-test is removed.
Dispatch-only fits the intermittent runner and keeps the 4-target Linux/
Windows release independent. Becomes triggerable once CI is on the default
branch (workflow_dispatch is default-branch-only in Gitea).
- Add `cargo test` before the darwin builds (gate is Linux-only; the macOS
leg is test-then-build) — a full dry-run of release-macos bar the upload.
- Add an `if: always()` prune step. The runner wipes the workspace each run,
so cargo target/ never accumulates (no sweep). The persistent cache is the
nix store: record the current toolchain in a persistent profile, keep the
2 newest generations (nix-env --delete-generations +2), reclaim the rest
(nix-collect-garbage). Pairs with min-free/max-free in the runner nix.conf.
libiconv is the only /nix/store dep the darwin stdenv bakes in (everything
else is system frameworks + libSystem/libobjc). The smoke-test now rewrites
that load path to /usr/lib/libiconv.2.dylib (ABI-compatible, present on
every Mac), re-signs ad-hoc (install_name_tool breaks the sig; arm64
requires a valid one), then verifies no /nix/store paths remain, the
signature is valid, and the native binary launches. Flake comment updated
to reflect the propagated-libiconv reality.
The smoke-test caught the aarch64 binary linking a /nix/store libiconv.dylib
— non-portable (won't exist on a user's Mac). The Apple SDK already provides
a system libiconv stub, so removing pkgs.libiconv makes the linker resolve
-liconv to /usr/lib instead. The smoke-test now fails if any /nix/store dylib
is linked.
Add the two *-apple-darwin targets to rust-toolchain.toml and apple-sdk +
libiconv to the flake devShell (darwin only) so the nix toolchain links
AppKit; make cargo-zigbuild/zig Linux-only (macOS builds natively). Repoint
the throwaway macOS workflow to actually build both darwin targets through
the flake on the Tart runner — the first real check of the macOS leg, which
can't be verified locally. Delete once release-macos lands.
In act_runner a label is `<name>:<backend>`; `:host` is the execution-
backend schema (run on host, no container), not part of the label. The
runner registered as `macos:host` therefore has the label `macos`, which
is what runs-on must reference.
Gitea only exposes workflow_dispatch for workflows on the default branch
(main); our CI is on `ci`, so the manual-run button/API isn't available.
Add a push trigger (filtered to the probe file) so we can drive the macOS
runner test from the ci branch. workflow_dispatch kept for post-merge.
Manual-dispatch probe on runs-on macos:host to confirm the runner picks up
jobs and report arch / macOS version / Xcode SDK / toolchains (nix, rustup,
cargo) / git+node, before wiring the macOS release leg. Delete once done.
Record the multi-platform build strategy as its own decision: cargo-zigbuild
for the four non-macOS targets, the static/standalone posture per platform,
the Windows synchronization stub, the test->build matrix workflow, and the
macOS deferral with its licensing rationale (the public CI image can't carry
the SDK). Shrinks the ci-001 amendment to a pointer; updates the index.
Runtime-verified by the user: Linux x86_64 + Windows aarch64 run correctly.
release.yaml becomes test (once, host) -> build (matrix) over the four
cargo-zigbuild targets; each matrix job uploads its binary + .sha256 to
the shared release (idempotent create-or-get). Records the expansion in
ADR-ci-001 (2026-06-13 amendment); macOS stays deferred.
Replace the single-target musl cc with cargo-zigbuild + zig in the flake
devShell — one universal cross cc/linker (incl. rusqlite's bundled SQLite
C) for all four non-macOS D1 targets, added to rust-toolchain.toml:
x86_64/aarch64-unknown-linux-musl (static, D2)
x86_64-pc-windows-gnu, aarch64-pc-windows-gnullvm (standalone .exe)
Windows links -lsynchronization (std WaitOnAddress), which rust-overlay's
toolchain and zig's mingw don't ship; the symbols are forwarded by
kernel32, so an empty stub libsynchronization.a (ci/winstub/, wired via
.cargo/config.toml for the windows targets only) satisfies the linker.
Verified: all four build; linux static; windows valid PE32+.
Add paths-ignore (docs/**, **/*.md) to the gate's push + pull_request
triggers so markdown/docs-only changes don't run a full clippy+test that
can't change the outcome. Mixed code+docs pushes still gate (not all
files are ignored); flake/toolchain changes are deliberately not ignored.
Also refresh a stale ADR-0049 -> ADR-ci-002 comment reference.
Records the CI/release pipeline as ADR-ci-001 and relocates the nix-flake
ADR from main's ADR-0049 to ADR-ci-002 (content unchanged, history note
added). Both live in docs/ci/adr/ with a README index — a dated,
ci-segmented namespace disjoint from main's integer ADR sequence, the
same split the website subproject uses to avoid cross-branch number
collisions. Drops the ADR-0049 entry from docs/adr/README.
ci-001 covers the runner model, the baked nix CI image, the clippy+test
gate, the static-musl release on tag, trigger hygiene, auth, and scope.
- Run cargo test before the build so a tag never publishes untested code.
- Pin shell: bash on the scripted steps; the runner defaults to dash,
which rejected `set -o pipefail` and failed run 22's package step.
- Swap `file` (absent in the slim image) for `ls -l`.
Tag pushes ignore paths: filters, so a release tag spuriously rebuilt
the unchanged CI image and re-ran the gate on a commit the branch push
already gated. Add branches: ['**'] to both push triggers — tag pushes
no longer fire them (release.yaml owns tags). Pushing commits + a tag
together still gates the commits via the branch push.
On a v* tag, builds the x86_64-unknown-linux-musl binary in the CI image
and publishes it (+ .sha256) to a Gitea release via the API and the
auto GITEA_TOKEN. x86_64 Linux only for now; rest of the D1 matrix and
D3 packaging layer on later. Correctness comes from the branch gate.
rust-toolchain.toml gains the x86_64-unknown-linux-musl target; the
flake devShell gains a musl cc (pkgsCross.musl64) + CC/linker env so a
`cargo build --target …-musl` compiles rusqlite's bundled SQLite C and
links fully static (D2: single static binary, no runtime deps). Cargo
release profile strips symbols (13MB -> 10MB). Verified locally: the
musl binary is static-pie, statically linked, stripped, runs standalone.
Job-container image the gate runs in. node:22-bookworm-slim satisfies
the act_runner contract (sleep/bash/node) far more cheaply than the
catthehacker images; single-user nix installed on top (pre-create /nix
+ empty build-users-group so it installs as root in a container) with
the flake's devShell pre-warmed — CI enters a ready 1.95.0 toolchain in
~1.4s. Verified by local build. ~5.5GB (rust toolchain closure); dep/
target caching is a noted follow-up.
Diagnostic to determine how the ci-public runner executes jobs and
where the nix toolchain is reachable (host vs default container vs a
custom container:), so the real gate is built on facts. Delete once
the gate lands.
Root flake with two outputs: devShells.default (pinned 1.95.0
toolchain via rust-toolchain.toml + rust-overlay, plus cargo-sweep)
and packages.default (rustPlatform.buildRustPackage from the committed
Cargo.lock; doCheck=false). flake.lock pins nixpkgs nixos-26.05 /
rust-overlay / flake-utils. .envrc (use flake) for direnv parity.
Single source of toolchain for dev and the upcoming CI, so they can't
drift. Verified through the flake: nix build yields a working binary,
clippy clean, 2424 tests pass / 0 fail / 1 intentional ignored doctest.
First step toward requirements.md TT5 + D1/D2/D3.
`decimal` is stored as exact TEXT, but SQLite has no native decimal type,
so arithmetic/aggregation implicitly coerces it to an IEEE-754 double.
The computed result carries no playground type, so `sum(price * qty)`
rendered the double's full noise — `298.59999999999997` for `298.60` — a
confusing, off-topic float lesson for a teaching tool.
Add `format_real_display`: round REAL values to 15 significant figures
(a double's reliable precision) then take the shortest round-tripping
form, collapsing `298.59999999999997` to `298.6`. Wired into `format_cell`
(result-set / `show data` cells) only — the sole surface where the noise
appears, since it arises from arithmetic.
Every other f64->string path keeps full precision for semantic, not
cosmetic, reasons: CSV persistence stays byte-exact for round-trip;
`render_value` is a canonical identity key for the uniqueness dry-runs
(dry_run_unique, check_uniqueness_collisions), where rounding would
report collisions the exact-valued engine wouldn't; FK-key matching and
EXPLAIN-SQL literals likewise stay exact.
ADR-0005 Amendment 1; +7 tests.
A bare table alias typed where a column is expected — `… GROUP BY o`,
with `o` aliasing `FROM Orders o` — was a blind spot: completion offered
nothing for `o`, and the hint panel called the in-scope alias an unknown
column (`no such column o on table Orders, ...`).
Completion now offers each FROM source's qualifier (alias-if-present-else
table-name) at a bare sql_expr_ident slot, folded into the column
candidate list; on an exact-qualifier partial the alias source steps
aside so the diagnostic can surface. The bare-reference diagnostic arm
emits a targeted `alias_used_as_column` / `table_used_as_column` hint
("`o` is a table alias — write `o.<column>` ...") after the
projection-alias check, so ORDER-BY alias refs still win and a genuine
unknown column still reports `unknown_column`.
Two guards keep the qualified-form advice correct: SQL only (role
`sql_expr_ident`, so the DSL `expr_column` path keeps `unknown_column`
since the DSL has no `table.column` syntax) and effective-qualifier
match (alias-if-present-else-table, so an aliased source referenced by
its shadowed real name falls through rather than being advised as
`name.<column>`). The diagnostic is a drop-in replacement for
`unknown_column` at the same span/Error severity, so verdict/overlay/hint
paths are unchanged.
ADR-0032 Amendment 3; +10 tests.
Captures this session's three committed fixes (flag-completion dash,
table-alias validity, `:` one-shot live feedback), the two
investigated-as-designed findings (comma-FROM, aggregate-without-GROUP
BY), and points the next session at the filed enhancement issues
(#26 seed-count hint, #27 bottom status line, #28 add-column
relationship prose) plus the standing roadmap.
The `:` one-shot escape (ADR-0003) is stripped at submission, but the
*live* feedback kept the leading `:` in the buffer it handed the
walker — so Tab completion, the validity verdict, and the highlight
overlays all bailed at the `:` and treated the SQL as an unknown
command. Effect: in `:`-mode, Tab completed nothing and a valid query
could flash an error, while the identical line in full `mode advanced`
worked. (The ambient hint already stripped it, which is why the hint
showed the right column name while Tab did nothing.)
Add `App::feedback_view()` — the `:`-stripped SQL view, the cursor
mapped into it, and the stripped byte offset — and route all four live
paths through it:
- completion (Tab): complete against the view, then shift the returned
`replaced_range` back by the offset so the edit lands in the buffer;
- validity verdict: verdict the SQL, not the sigil;
- highlight/overlays: new `render_input_runs_feedback` highlights and
diagnoses the view (shifted by the offset) while the `:` renders as
plain text;
- ambient hint: consolidated onto `feedback_view`, replacing the
duplicate local `strip_one_shot_prefix`.
In a SELECT, the projection can reference a table alias whose defining
FROM binding sits textually *after* the cursor — e.g.
`select sum(ol.count*p.price) … from … OrderLines ol …`. The candidate
engine already recovers that scope via the §10.6 full-input lookahead
(ADR-0032), but the typing-time validity indicator
(`invalid_ident_at_cursor`) walked only the text before the cursor,
found `ol` in no scope, and flagged it as an unknown column — a red
"ERR" overlay on an otherwise-valid query. (Other aliases escaped only
by coincidentally prefix-matching real columns.)
Give the validity check the same full-input lookahead: at a SQL
expression slot, recover the from-scope from the whole input and bail
when the partial prefix-matches a binding's alias or table name.
The partial-token walk stopped at `-`, so after typing `-` (or `--`)
the partial was empty and the replaced range was a zero-width point
*after* the dash. Two bugs followed at a flag position (e.g.
`add 1:n relationship … -`): the `on` keyword was offered (it
prefix-matched the empty partial), and accepting a candidate inserted
after the dash — `-on`, `---create-fk`, `----all-rows`.
Detect a dash-prefixed token at a word boundary as a flag-in-progress
and fold the whole dash-run into the partial, gated on a flag actually
being expected there (so `where x = -5` stays a signed number, not a
flag). The flag matcher now strips leading dashes and matches the body
uniformly (empty / `-` / `--` → all flags; `--cr` → create-fk).
Keywords like `on` no longer appear after a dash, and accept replaces
the dash(es) so `-` → `--create-fk` and `--all` → `--all-rows`. Two
partial-flag snapshots updated (they had captured the old behaviour).
Build the two SD2 surfaces Phase 1 deferred:
- `set` override clause (D2): comma-separated per-column pins —
`= 'v'` (fixed), `in ('a','b')` (pick-list), `as <generator>`
(named), `between x and y` (range; numeric and quoted dates).
Type-aware via the typed `current_column_value` slot; an override
drops its column from the generic-fill advisory (D13). Folded from
the flat matched path (build_seed_overrides) and applied to the
per-column plan (apply_seed_overrides).
- `<table>.<column>` column-fill (D1 form 2): an UPDATE over existing
rows. Refuses PK/autogen targets, empty-table no-op, FK-samples the
parent, collision-free for UNIQUE/identifier targets, one undo step;
`set` may only adjust the filled column.
Supporting work: KNOWN_GENERATORS vocabulary + generator_for_name
(src/seed/vocabulary.rs, D9); a range Generator + range_bounds_reason;
IdentSource::Generators and HighlightClass::Function; completion of the
generator vocabulary after `as` and the set/.col column slots; the
typing-time validity indicator for an unknown generator; help,
parse-error pedagogy rows, and the D13 advisory's Phase-2/3 wording.
A bounded override (fixed value / too-short pick-list) on a
single-column-UNIQUE target is a friendly error rather than a silent
uniqueness cap (post-implementation /runda finding, user-chosen).
Dates in the range form are quoted (no date-literal token exists);
ADR-0048 D2 amended accordingly. Both modes (D5); reproducible (D4).
- ADR-0048 status -> Accepted; Phase 1 implemented (commits
202e25a..fbd219b), with the pre-build and post-implementation /runda
passes and the 2358-test green state recorded; index entry updated.
- requirements.md: SD1 [x] (whole-row seed + FK/junction, both modes,
--seed reproducibility with no exceptions), SD2 [/] (core generators /
determinism done; the set override clause + column-fill are Phase 2),
A1 14/15 (only hint/H2 remains unregistered).
- Handoff 65: the full seed Phase-1 build, the two /runda passes, where
the code lives, and Phase-2 / next steps.
P1.4 — user-visible surface:
- Grammar: `seed <table> [count] [--seed <n>]` (the first DSL flag with a
value); build_seed disambiguates the seed value from the positional count.
- Verified the auto-wired surface: table-name completion, --seed offered as
a candidate, validity consistent with `show data`, an ADR-0042 near-miss
row for bare `seed`, and render tests for the seed outcome.
/runda hardening — eight DA findings, all resolved:
- FK sampling now uses ORDER BY so --seed reproducibility no longer relies
on SQLite's unspecified DISTINCT order (D4).
- shortid columns now generate from seed's seeded RNG (new
shortid::generate_with_rng) — D4 now holds with no exceptions.
- Added the missing coverage the DA flagged: undo-one-step (D15), replay
re-runs a seed line (D16), advanced-mode (D5), atomic rollback on a
constraint failure, seed 0 no-op, complex-CHECK advisory (D17), and
FK + shortid reproducibility.
2358 pass / 0 fail / 0 skip, clippy all-targets clean.
do_seed inserted row-by-row through do_insert, re-writing the whole
table CSV each time — O(N^2). Extract do_insert's row core into a new
insert_one_row (bind + serial/shortid autofill + FK-enriched execute,
no tx/persist), shared by:
- do_insert: one row in its own transaction (behaviour unchanged).
- do_seed: all rows in ONE transaction, with a single
finalize_persistence before the single commit — O(N), preserving
ADR-0015 §6 commit-db-last. A mid-batch failure now rolls the whole
seed back atomically; the capped preview is read back by rowid.
A near-max 10000-row seed drops from ~tens of seconds to well under
one. do_insert behaviour unchanged (whole suite green: 2346 pass /
0 fail / 0 skip, clippy clean); seed's existing tests exercise the
batch path.
A dedicated SeedResult replaces the borrowed insert outcome (X5):
- CommandOutcome::Seed + DslSeedSucceeded event + handle_dsl_seed_success
render: the echo, "N row(s) seeded into T", a capped preview table
(D18, first 20 rows; full count always reported), and a Hint-styled
advisory naming enum-ish / un-derivable-CHECK columns filled with
generic text (D12/D13, Phase-1 wording).
- SeedResult carries requested vs produced, so a junction cap is now
reported to the user, not only logged.
- Count cap (D6): a seed over 10000 rows is refused with a friendly error.
- Catalog keys ok.rows_seeded / seed.capped / seed.advisory_generic.
4 new tests (advisory flag, IN-check not flagged, preview cap, excess
count). 2346 pass / 0 fail / 0 skip, clippy clean.
do_seed now enforces value uniqueness and derives enum values:
- Uniqueness groups (D10): the user-fillable PK, compound UNIQUE
constraints, and single-column UNIQUE / identifier columns stay
distinct across the batch and against existing rows (retry per row).
Junction distinct-combos fall out of PK-tuple uniqueness and cap at
the available parent combinations (logged when capped; the
user-facing note arrives with the advisory in P1.3c).
- Identifier-int columns get a monotonic sequence past MAX(col) (D10),
so they never collide.
- IN-CHECK derivation (D17): a simple `col IN ('a','b')` CHECK becomes
the value source via the new, unit-tested seed::parse_in_check_values,
so the enum-as-CHECK pattern just works.
8 parser unit tests + 4 integration tests (unique column, identifier
sequencing, junction cap, IN-check enum). 2343 pass / 0 fail / 0 skip,
clippy all-targets clean.
Deferred to P1.3c: dedicated SeedResult + capped preview (D18) + the
enum/CHECK advisory incl. the cap note (D12/D13); P1.3d: multi-row path.
do_seed fills foreign-key columns by sampling existing parent rows
(D14): sample_parent_key_tuples reads distinct parent keys, and a
compound FK reads all its child columns from one sampled parent row per
child row. An empty parent is refused with a friendly "seed the parent
first" error. The block guard (D1) refuses a NOT NULL blob column (seed
can't generate one); a nullable blob is omitted (-> NULL).
4 integration tests (valid FK references, empty-parent refusal, NOT NULL
blob refusal, nullable-blob omission). 2331 pass / 0 fail / 0 skip,
clippy all-targets clean.
Deferred to P1.3b: identifier/constraint uniqueness incl. junction
distinct-combos (D10), IN-CHECK derivation (D17), dedicated SeedResult +
capped preview (D18) + advisory (D12/D13), and the multi-row path.
The pure generation half of `seed` — no command wiring yet:
- src/seed/: ColumnSpec + Generator model and a seeded StdRng; the
type-gated name-heuristic catalogue (D7) with documented
false-positive guards; table-context name disambiguation (D11);
identifier (D10) and enum-ish (D12) detection; per-type + bounded-date
generators (D8); the hand-rolled product generator (D9); and PickFrom
for IN-CHECK / enum lists.
- Adds the `fake` crate (v5, default features). Verified: single rand
0.10.1 (no duplication), determinism via one seeded StdRng driving
both fake and the hand-rolled generators, security-clean across
osv/grype/trivy.
- ADR-0048 D3 updated to record the dependency verification.
32 Tier-1 tests (exact-value via fixed --seed); 1673 lib tests pass,
clippy all-targets clean.
Dedicated `seed` command: realistic, name-aware fake data via the
`fake` crate + a type-gated heuristic catalogue, table-context name
disambiguation, a hand-rolled product generator, bounded dates, an
identifier-uniqueness rule, a quoted `set` override clause (value /
list / generator / range), `--seed` reproducibility, FK sampling
(empty-parent error, junction distinct combinations), CHECK
derive-or-friendly-fail, undo as one batch step, replay as a
data-write, capped auto-show, and an enum/CHECK advisory.
Design settled across an extended user fork dialogue, then hardened
by a /runda DA pass that found six blockers (undo, replay, set
quoting, CHECK handling, advisory phasing, auto-show flood) — all
folded in, genuine forks re-escalated and user-resolved.
Takes up SD2 ([~]->[ ]); SD1 implementation in progress.
The trailing "multiple tabs" clause of S3 and V2 is dropped from
tracked scope by user decision. The output pane is settling on the
single scrollable V4 journal model rather than switchable result
tabs, so both items are now fully satisfied (the table view and
single-element visualisation were already built). V4 noted as the
sole tracked direction for evolving the output pane. A future
return to tabbed output would be a fresh requirement.
No ADR touched: multi-tab output never had a deciding ADR.
Bring handoff-64 current: it was written just before #22/#24 were closed
and the docs commit landed. §1 now reflects HEAD f0afec3 with both
issues closed on Gitea; §6 removes the completed finalization steps and
points the next session at the open requirements backlog (§7).
Render the keystroke badge and step caption as a solid yellow rectangle
with no border glyphs and a one-cell text margin, instead of a
rounded-border box — deliberately unlike the app's bordered panels so
the demo overlays read as a distinct, eye-catching callout. Shared
fill_overlay_rect helper (borderless Block fill + inset Paragraph).
Snapshots regenerated; ADR-0047 D4 wording updated.
Ctrl+] (decodes to Char('5')+CONTROL) toggles an invisible capture
buffer: typed characters accumulate without touching the input/output,
Backspace edits, every other key is inert, and a second Ctrl+] commits
the text to a caption box (empty commit dismisses). Handled at the top
of handle_key — before the badge and modal gates — so captions can be
authored over the load picker (the #24 cast); an ordinary keystroke
clears a visible caption. The caption renders as a floating
black-on-yellow box at the output panel's bottom-right, wrapped to <=3
lines (then ellipsised), with the keystroke badge stacked directly
above it when both are present.
Tier 1: capture/commit, invisible accumulation, backspace, inert keys
(incl. no badge), empty-commit dismiss, next-key clear, over-modal,
demo-off inert. Tier 2: caption / stacked / wrapped snapshots. Phase C
of ADR-0047 — feature complete.
In --demo mode, an otherwise-invisible key (Tab, Enter, arrows,
Ctrl-O, …) raises a transient [LABEL] badge — a floating
black-on-yellow box inset at the output panel's bottom-right. Set in
App::update before the modal gate (so it shows over the load picker,
the #24 cast); pure demo_badge_label maps the key set. The runtime
expires it on a ~1.5s timer via a new nearest_deadline helper that
extends the existing time-boxed-recv arm condition without disturbing
the ADR-0027 indicator debounce. New App.last_output_area lets the
top-level draw anchor the overlay; overlay colours centralised in
theme.rs.
Tier 1 (label fn, badge set/seq, over-modal), Tier 2 (dark/light
snapshots, black-on-yellow style, too-small clamp), runtime unit
(nearest_deadline). Phase B of ADR-0047; captions land in C.
Add `--demo` (and the RDBMS_PLAYGROUND_DEMO env fallback) to enter
demonstration mode, threaded onto App.demo_mode through run_loop —
mirrors the --no-undo plumbing. Off by default, zero footprint when
off. The --help line advertises only the visible keystroke badges;
the Ctrl+] caption trigger is kept low-profile (ADR-0047 D6 updated).
Phase A of ADR-0047; behaviour (badges/captions) lands in B and C.
Accepted decision record for the in-app demo overlay: a --demo mode
that shows automatic keystroke badges ([TAB], [ENTER], …) and a
stealth Ctrl+]-delimited step-caption buffer, both as floating
black-on-yellow boxes at the output panel's bottom-right. All forks
user-confirmed; a /runda pass contributed 10 tightening findings.
Indexed in docs/adr/README.md.
Add j (down), k (up), g (first) and G (last) to the load picker's
list sub-mode, alongside the existing arrow keys. Typeable keys keep
the picker drivable by autocast in the website's documentation casts,
which cannot emit arrow keys. Footer hint left unchanged.
ADR-0046 status -> Accepted + implemented (8 commits 9f5f76b..22bec61);
README index updated; the two draft-divergent decisions recorded inline
(App.relationships not SchemaCache; nav overlay partial-clear + gutter).
Handoff 63 covers the full UI build across Phases A/B/C; issues
#20/#21/#23 closed on Gitea.
DC3 — navigation-mode scroll: the focused Tables / Relationships panel
scrolls (Up/Down by a line, PageUp/PageDown by its visible-row count).
Per-panel offsets are clamped to content at render time, and the
renderer reports each panel's visible rows for paging — mirroring the
output panel's scroll. render_items_panel / render_relationships_panel
take &mut App, count their rows, and store+clamp the offset before
building the borrowing lines.
DC2 refinement: the expand-on-focus overlay now clears only the sidebar
strip plus a one-column gutter, leaving the base output/input/hint
visible (unchanged) to the right rather than blanking the whole area —
truer to "underneath keeps its layout", with the gutter keeping the
cut-off edge clean (chosen after eyeballing both variants). ADR DC2 and
the overlay snapshot updated to match.
Tests: line/page scroll move only the focused panel and clamp; the
render clamps a past-the-end offset so the last row stays visible.
claude@clouddev1