delint/get basic async checks working

.mise.toml

@@ -2,4 +2,4 @@
 prek = "0.4.1"
 zig = "0.16.0"
 zls = "0.16.0"
-"ubi:DonIsaac/zlint" = "0.7.9"
+"ubi:DonIsaac/zlint" = "0.8.1"

.pre-commit-config.yaml

@@ -23,14 +23,6 @@ repos:
         entry: bash -c 'printf "%s\n" "$@" | zlint --deny-warnings --fix -S' --
         language: system
         types: [zig]
-        # zlint 0.7.9 misparses Zig 0.16's `async`/`await` method
-        # names (Group.async, Future.await) as the legacy keywords,
-        # firing false-positive "syntax error" diagnostics. Until
-        # zlint catches up, exclude files that legitimately use the
-        # std.Io.Group async/await methods. `zig fmt` rewrites the
-        # `@"async"`/`@"await"` workaround back to the bare form,
-        # so we can't dodge it locally either.
-        exclude: ^src/(tui|PortfolioData|service)\.zig$
   - repo: https://github.com/batmac/pre-commit-zig
     rev: v0.3.0
     hooks:

src/PortfolioData.zig

@@ -68,6 +68,7 @@
 //! `deinit()` cancels first, then frees.
 
 const std = @import("std");
+const log = std.log.scoped(.portfolio_data);
 
 const PortfolioData = @This();
 
@@ -517,21 +518,21 @@ pub fn load(
     defer watchlist_set.deinit();
     var portfolio_set = std.StringHashMap(void).init(gpa);
     defer portfolio_set.deinit();
-    for (syms) |s| portfolio_set.put(s, {}) catch {};
+    for (syms) |s| portfolio_set.put(s, {}) catch return error.OutOfMemory;
     for (opts.watchlist_syms) |sym| {
-        if (!portfolio_set.contains(sym)) watchlist_set.put(sym, {}) catch {};
+        if (!portfolio_set.contains(sym)) watchlist_set.put(sym, {}) catch return error.OutOfMemory;
     }
     for (pf.lots) |lot| {
         if (lot.security_type == .watch) {
             const sym = lot.priceSymbol();
-            if (!portfolio_set.contains(sym)) watchlist_set.put(sym, {}) catch {};
+            if (!portfolio_set.contains(sym)) watchlist_set.put(sym, {}) catch return error.OutOfMemory;
         }
     }
     var watch_syms_list: std.ArrayList([]const u8) = .empty;
     defer watch_syms_list.deinit(gpa);
     {
         var it = watchlist_set.keyIterator();
-        while (it.next()) |k| watch_syms_list.append(gpa, k.*) catch {};
+        while (it.next()) |k| watch_syms_list.append(gpa, k.*) catch return error.OutOfMemory;
     }
 
     // ── Fetch prices ──────────────────────────────────────────
@@ -579,10 +580,16 @@ pub fn load(
         const k = entry.key_ptr.*;
         const v = entry.value_ptr.*;
         if (portfolio_set.contains(k)) {
-            prices.put(k, v) catch {};
+            prices.put(k, v) catch return error.OutOfMemory;
         } else if (watchlist_set.contains(k)) {
             const owned = arena_alloc.dupe(u8, k) catch continue;
-            wp.put(owned, v) catch {};
+            wp.put(owned, v) catch |err| {
+                // Free the just-duped key... except it's
+                // arena-allocated, so the arena reaps it.
+                // Skip this watchlist price; a missing
+                // watchlist entry degrades display only.
+                log.warn("watchlist price put({s}): {t}", .{ k, err });
+            };
         }
     }
     self.latest_quote_date = load_all.latest_date;

src/analytics/observations.zig

@@ -104,26 +104,52 @@ pub const Check = struct {
     run: *const fn (ctx: CheckCtx) CheckResult,
 };
 
-/// Per-check execution state, lives on `CheckPanel`. Today every
-/// check is sync so `state` is always `.complete` immediately after
-/// `runChecks` returns. The async path is in place for future use.
+/// Per-check execution state, lives on `CheckPanel`. Sync checks
+/// are `.complete` immediately after `runChecks` returns;
+/// `is_long_running` checks start `.pending` and transition to
+/// `.complete` via `poll` or `awaitResult`.
 pub const PendingCheck = struct {
     check: *const Check,
+    /// Completion flag set by the async wrapper as its final
+    /// action. `poll` reads this to detect completion without
+    /// blocking — `std.Io.Future` itself only offers blocking
+    /// `await`/`cancel`, so the flag is what makes a
+    /// non-blocking probe possible. The tiny window between
+    /// flag-set and the future's internal result write is
+    /// covered by the (then nearly instant) `await` in `poll`.
+    done: std.atomic.Value(bool) = .init(false),
     state: union(enum) {
         complete: CheckResult,
-        // Future: pending: std.Io.Future(CheckResult), populated
-        // by io.async() and resolved via poll/await. Adding a
-        // variant here is a non-breaking API change for current
-        // consumers.
+        pending: std.Io.Future(CheckResult),
     },
 
-    /// Returns the resolved `CheckResult`, awaiting if necessary.
-    /// Idempotent.
+    /// Returns the resolved `CheckResult`, awaiting (blocking) if
+    /// necessary. Idempotent.
     pub fn awaitResult(self: *PendingCheck, io: std.Io) CheckResult {
-        _ = io;
-        return switch (self.state) {
-            .complete => |r| r,
-        };
+        switch (self.state) {
+            .complete => |r| return r,
+            .pending => |*f| {
+                const r = f.await(io);
+                self.state = .{ .complete = r };
+                return r;
+            },
+        }
+    }
+
+    /// Non-blocking completion probe. Returns true when the
+    /// result is available in `.complete` (transitioning it
+    /// there if the async task just finished). Returns false
+    /// while the task is still running.
+    pub fn poll(self: *PendingCheck, io: std.Io) bool {
+        switch (self.state) {
+            .complete => return true,
+            .pending => |*f| {
+                if (!self.done.load(.acquire)) return false;
+                const r = f.await(io);
+                self.state = .{ .complete = r };
+                return true;
+            },
+        }
     }
 };
 
@@ -132,23 +158,35 @@ pub const PendingCheck = struct {
 /// per-finding strings inside each `CheckResult`).
 pub const CheckPanel = struct {
     allocator: std.mem.Allocator,
+    io: std.Io,
     pending: []PendingCheck,
 
     pub fn deinit(self: *CheckPanel) void {
-        for (self.pending) |pc| freeResult(self.allocator, pc.state.complete);
+        for (self.pending) |*pc| {
+            // Resolve any still-pending future before freeing.
+            // `cancel` requests cancelation and blocks until the
+            // task returns — the task may still produce a full
+            // result (checks don't hit cancelation points today),
+            // which we then free like any complete result.
+            const result = switch (pc.state) {
+                .complete => |r| r,
+                .pending => |*f| f.cancel(self.io),
+            };
+            freeResult(self.allocator, result);
+        }
         self.allocator.free(self.pending);
         self.* = undefined;
     }
 
-    /// True iff every check has a resolved result. Today always
-    /// returns `true` immediately after `runChecks` (sync only).
-    pub fn isComplete(self: *const CheckPanel) bool {
-        for (self.pending) |pc| {
-            switch (pc.state) {
-                .complete => {},
-            }
+    /// True iff every check has a resolved result. Non-blocking:
+    /// polls each pending check, transitioning newly-finished
+    /// ones to `.complete` as a side effect.
+    pub fn isComplete(self: *CheckPanel) bool {
+        var all = true;
+        for (self.pending) |*pc| {
+            if (!pc.poll(self.io)) all = false;
         }
-        return true;
+        return all;
     }
 };
 
@@ -175,25 +213,66 @@ fn freeObservations(a: std.mem.Allocator, obs: []const Observation) void {
 
 // ── Runner ────────────────────────────────────────────────────
 
+/// Async wrapper for long-running checks. Runs the check, then
+/// sets the completion flag as the final action so `poll` can
+/// detect the result without blocking.
+///
+/// IMPORTANT: `done` points into the panel's `pending` array,
+/// and `ctx`'s borrowed slices (`rows`, `totals`) must outlive
+/// the task. Both invariants are owned by `runChecks`'s caller
+/// contract: the panel and the review view that owns ctx's rows
+/// live until `CheckPanel.deinit`, which resolves all futures
+/// before anything is freed.
+fn runCheckTask(check: *const Check, ctx: CheckCtx, done: *std.atomic.Value(bool)) CheckResult {
+    const result = check.run(ctx);
+    done.store(true, .release);
+    return result;
+}
+
 /// Run the registered checks against the given context. Returns a
-/// `CheckPanel` that renderers consume. Sync today; the contract
-/// allows a future async dispatch path for `is_long_running` checks.
+/// `CheckPanel` that renderers consume.
+///
+/// Sync checks (`is_long_running == false`) complete inline; the
+/// panel entry is `.complete` on return. Long-running checks are
+/// dispatched via `io.async` and start `.pending`; callers
+/// resolve them with `PendingCheck.poll` (non-blocking, for
+/// progressive TUI rendering) or `awaitResult` (blocking, for
+/// CLI output).
+///
+/// Lifetime contract: `ctx.rows` / `ctx.totals` are borrowed by
+/// in-flight async checks. The caller must keep them alive until
+/// every pending check resolves — in practice both the panel and
+/// the rows live on the same `ReviewView` and are torn down
+/// together by `ReviewView.deinit` (panel first, which resolves
+/// stragglers via cancel).
 pub fn runChecks(
     allocator: std.mem.Allocator,
     io: std.Io,
     ctx: CheckCtx,
     checks: []const Check,
 ) !CheckPanel {
-    _ = io;
     var pending = try allocator.alloc(PendingCheck, checks.len);
     errdefer allocator.free(pending);
 
+    // Two passes: initialize every slot first so the `done`
+    // flags have stable addresses, THEN spawn the async tasks
+    // that point at them. (Spawning during the init loop would
+    // hand out pointers into an array we're still writing.)
     for (checks, 0..) |*check, i| {
-        const result = check.run(ctx);
-        pending[i] = .{ .check = check, .state = .{ .complete = result } };
+        pending[i] = .{ .check = check, .state = .{ .complete = .skipped } };
     }
 
-    return .{ .allocator = allocator, .pending = pending };
+    for (checks, 0..) |*check, i| {
+        if (check.is_long_running) {
+            pending[i].state = .{
+                .pending = io.async(runCheckTask, .{ check, ctx, &pending[i].done }),
+            };
+        } else {
+            pending[i].state = .{ .complete = check.run(ctx) };
+        }
+    }
+
+    return .{ .allocator = allocator, .io = io, .pending = pending };
 }
 
 // ── Threshold constants ───────────────────────────────────────
@@ -1001,6 +1080,9 @@ test "runChecks: produces a panel with one entry per check" {
     var panel = try runChecks(testing.allocator, std.testing.io, ctx, &default_checks);
     defer panel.deinit();
     try testing.expectEqual(default_checks.len, panel.pending.len);
+    // Await everything (drift is is_long_running and runs async);
+    // after awaiting, the panel must report complete.
+    for (panel.pending) |*pc| _ = pc.awaitResult(std.testing.io);
     try testing.expect(panel.isComplete());
 }
 
@@ -1012,12 +1094,82 @@ test "runChecks: empty portfolio every check passes or skips" {
     };
     var panel = try runChecks(testing.allocator, std.testing.io, ctx, &default_checks);
     defer panel.deinit();
-    for (panel.pending) |pc| {
-        const tag = std.meta.activeTag(pc.state.complete);
+    for (panel.pending) |*pc| {
+        const tag = std.meta.activeTag(pc.awaitResult(std.testing.io));
         try testing.expect(tag == .pass or tag == .skipped);
     }
 }
 
+test "runChecks: async check resolves via poll without blocking forever" {
+    // Pin the async-dispatch contract: an is_long_running check
+    // starts pending and becomes complete via poll once its task
+    // finishes. We can't observe the intermediate pending state
+    // deterministically (the task may finish before we look),
+    // but we CAN assert the poll loop terminates and yields the
+    // right result.
+    const slow_check = [_]Check{.{
+        .name = "test_async",
+        .label = "Test async",
+        .is_long_running = true,
+        .run = struct {
+            fn run(c: CheckCtx) CheckResult {
+                _ = c;
+                return .pass;
+            }
+        }.run,
+    }};
+    const ctx: CheckCtx = .{
+        .allocator = testing.allocator,
+        .rows = &.{},
+        .totals = emptyTotals(),
+    };
+    var panel = try runChecks(testing.allocator, std.testing.io, ctx, &slow_check);
+    defer panel.deinit();
+
+    // Poll until complete (bounded loop; the task is trivially
+    // fast, so thousands of iterations would indicate a real
+    // hang — fail rather than spin forever).
+    var iterations: usize = 0;
+    while (!panel.isComplete()) : (iterations += 1) {
+        try testing.expect(iterations < 1_000_000);
+    }
+    try testing.expectEqual(CheckResult.pass, panel.pending[0].awaitResult(std.testing.io));
+}
+
+test "runChecks: deinit with unresolved async check does not leak or crash" {
+    // The panel may be torn down while a check is still pending
+    // (user quits the TUI mid-poll). deinit must cancel/resolve
+    // the future and free whatever result it produced.
+    const slow_check = [_]Check{.{
+        .name = "test_async_abandon",
+        .label = "Test abandon",
+        .is_long_running = true,
+        .run = struct {
+            fn run(c: CheckCtx) CheckResult {
+                // Allocate a real finding so deinit has something
+                // to free — exercises the result-ownership path.
+                const obs = c.allocator.alloc(Observation, 1) catch return .pass;
+                obs[0] = .{
+                    .severity = .warn,
+                    .kind = c.allocator.dupe(u8, "test_async_abandon") catch return .pass,
+                    .target = c.allocator.dupe(u8, "TEST") catch return .pass,
+                    .text = c.allocator.dupe(u8, "test finding") catch return .pass,
+                };
+                return .{ .warn = obs };
+            }
+        }.run,
+    }};
+    const ctx: CheckCtx = .{
+        .allocator = testing.allocator,
+        .rows = &.{},
+        .totals = emptyTotals(),
+    };
+    var panel = try runChecks(testing.allocator, std.testing.io, ctx, &slow_check);
+    // Deinit immediately — no poll, no await. testing.allocator
+    // catches any leak of the result allocations.
+    panel.deinit();
+}
+
 test "default_checks: every check name is unique" {
     for (default_checks, 0..) |a, i| {
         for (default_checks[i + 1 ..]) |b| {

src/atomic.zig

@@ -46,7 +46,12 @@ pub fn writeFileAtomic(
         });
         errdefer {
             tmp_file.close(io);
-            std.Io.Dir.cwd().deleteFile(io, tmp_path) catch {};
+            // Best-effort cleanup of the temp file while unwinding
+            // the primary error; a failed delete leaves a stray
+            // .tmp that the next write overwrites.
+            std.Io.Dir.cwd().deleteFile(io, tmp_path) catch |err| {
+                std.log.debug("atomic write cleanup deleteFile({s}): {t}", .{ tmp_path, err });
+            };
         }
 
         try tmp_file.writeStreamingAll(io, bytes);
@@ -58,7 +63,11 @@ pub fn writeFileAtomic(
     }
 
     std.Io.Dir.cwd().rename(tmp_path, std.Io.Dir.cwd(), path, io) catch |err| {
-        std.Io.Dir.cwd().deleteFile(io, tmp_path) catch {};
+        // Same best-effort cleanup as above; the rename failure is
+        // the error the caller needs to see.
+        std.Io.Dir.cwd().deleteFile(io, tmp_path) catch |del_err| {
+            std.log.debug("atomic write cleanup deleteFile({s}): {t}", .{ tmp_path, del_err });
+        };
         return err;
     };
 }

src/commands/review.zig

@@ -222,6 +222,14 @@ pub fn run(ctx: *framework.RunCtx, parsed: ParsedArgs) !void {
     );
     defer view.deinit(allocator);
 
+    // CLI is one-shot output — block until every async check
+    // resolves so the rendered grid + findings are complete.
+    // (The TUI renders progressively instead; see review_tab's
+    // tick hook.)
+    if (view.observations) |*panel| {
+        for (panel.pending) |*pc| _ = pc.awaitResult(io);
+    }
+
     // Sort: explicit --sort overrides the default grouping.
     if (parsed.sort) |field| {
         review_view.sortRows(view.rows, field, parsed.sort_dir);
@@ -359,7 +367,13 @@ fn renderStatusGrid(
         var worst: u8 = 0; // 0=pass/skipped, 1=warn, 2=flag/err
         var worst_color: [3]u8 = cli.CLR_MUTED;
         for (panel.pending[i..end]) |pc| {
-            const result = pc.state.complete;
+            // The CLI awaits every check before rendering (see
+            // run()), so .pending here would be a logic bug —
+            // skip defensively rather than crash.
+            const result = switch (pc.state) {
+                .complete => |r| r,
+                .pending => continue,
+            };
             const rank: u8 = switch (result) {
                 .pass, .skipped => 0,
                 .warn => 1,
@@ -390,12 +404,15 @@ fn renderStatusGrid(
             }
             try out.print("{s} ", .{label});
 
-            const glyph: []const u8 = switch (pc.state.complete) {
-                .pass => "✅\u{FE0F}",
-                .warn => "⚠️",
-                .flag => "❌\u{FE0F}",
-                .skipped => "➖\u{FE0F}",
-                .err => "🛑\u{FE0F}",
+            const glyph: []const u8 = switch (pc.state) {
+                .complete => |r| switch (r) {
+                    .pass => "✅\u{FE0F}",
+                    .warn => "⚠️",
+                    .flag => "❌\u{FE0F}",
+                    .skipped => "➖\u{FE0F}",
+                    .err => "🛑\u{FE0F}",
+                },
+                .pending => "⏳\u{FE0F}", // see defensive note above
             };
             try out.print("{s}", .{glyph});
         }

src/net/RateLimiter.zig

@@ -64,9 +64,14 @@ pub fn tryAcquire(self: *RateLimiter) bool {
 /// Acquire a token, blocking (sleeping) until one is available.
 pub fn acquire(self: *RateLimiter) void {
     while (!self.tryAcquire()) {
-        // Sleep for the time needed to generate 1 token
+        // Sleep for the time needed to generate 1 token. An
+        // interrupted sleep (cancelation propagating through the
+        // Io) just loops back to tryAcquire — the next refill
+        // covers whatever fraction of the wait elapsed.
         const wait_ns: u64 = @intFromFloat(1.0 / self.refill_rate_per_ns);
-        std.Io.sleep(self.io, .{ .nanoseconds = @intCast(wait_ns) }, .awake) catch {};
+        std.Io.sleep(self.io, .{ .nanoseconds = @intCast(wait_ns) }, .awake) catch |err| {
+            std.log.scoped(.rate_limiter).debug("acquire sleep interrupted: {t}", .{err});
+        };
     }
 }
 
@@ -74,7 +79,11 @@ pub fn acquire(self: *RateLimiter) void {
 /// Use after receiving a server-side 429 to wait before retrying.
 pub fn backoff(self: *RateLimiter) void {
     const wait_ns: u64 = @max(self.estimateWaitNs(), 2 * std.time.ns_per_s);
-    std.Io.sleep(self.io, .{ .nanoseconds = @intCast(wait_ns) }, .awake) catch {};
+    // Interrupted backoff sleep degrades to a shorter wait; the
+    // caller's retry may hit 429 again and re-backoff.
+    std.Io.sleep(self.io, .{ .nanoseconds = @intCast(wait_ns) }, .awake) catch |err| {
+        std.log.scoped(.rate_limiter).debug("backoff sleep interrupted: {t}", .{err});
+    };
 }
 
 /// Returns estimated wait time in nanoseconds until a token is available.

src/padded.zig

@@ -28,9 +28,11 @@ pub fn Padded(comptime T: type) type {
             // we can measure its length and pad. 64 bytes covers
             // every realistic format-method output in this codebase
             // (Money's worst case is ~14 chars; Date is exactly 10).
+            // An inner value exceeding the buffer surfaces as
+            // WriteFailed to the caller rather than UB.
             var tmp: [64]u8 = undefined;
             var fixed = std.Io.Writer.fixed(&tmp);
-            self.inner.format(&fixed) catch unreachable;
+            try self.inner.format(&fixed);
             const text = fixed.buffered();
 
             const pad = if (text.len >= self.width) 0 else self.width - text.len;

src/portfolio_loader.zig

@@ -742,7 +742,7 @@ fn gitInTestRepo(allocator: std.mem.Allocator, cwd: []const u8, argv: []const []
     defer allocator.free(result.stderr);
     switch (result.term) {
         .exited => |code| if (code != 0) {
-            std.debug.print("git command failed (code {d}): {s}\nstderr: {s}\n", .{ code, std.mem.join(allocator, " ", argv) catch "?", result.stderr });
+            std.log.err("git command failed (code {d}): {s}\nstderr: {s}", .{ code, std.mem.join(allocator, " ", argv) catch "?", result.stderr });
             return error.GitFailed;
         },
         else => return error.GitFailed,

src/providers/tiingo.zig

@@ -96,8 +96,11 @@ pub const Tiingo = struct {
     ) !CandleAndCorporateActions {
         var from_buf: [10]u8 = undefined;
         var to_buf: [10]u8 = undefined;
-        const from_str = std.fmt.bufPrint(&from_buf, "{f}", .{from}) catch unreachable;
-        const to_str = std.fmt.bufPrint(&to_buf, "{f}", .{to}) catch unreachable;
+        // Date's `{f}` output is exactly 10 bytes (YYYY-MM-DD), so
+        // these cannot fail in practice; `try` instead of
+        // `catch unreachable` to keep the error path honest.
+        const from_str = try std.fmt.bufPrint(&from_buf, "{f}", .{from});
+        const to_str = try std.fmt.bufPrint(&to_buf, "{f}", .{to});
 
         const symbol_url = try std.fmt.allocPrint(allocator, base_url ++ "/{s}/prices", .{symbol});
         defer allocator.free(symbol_url);

src/providers/twelvedata.zig

@@ -52,8 +52,11 @@ pub const TwelveData = struct {
 
         var from_buf: [10]u8 = undefined;
         var to_buf: [10]u8 = undefined;
-        const from_str = std.fmt.bufPrint(&from_buf, "{f}", .{from}) catch unreachable;
-        const to_str = std.fmt.bufPrint(&to_buf, "{f}", .{to}) catch unreachable;
+        // Date's `{f}` output is exactly 10 bytes (YYYY-MM-DD), so
+        // these cannot fail in practice; `try` instead of
+        // `catch unreachable` to keep the error path honest.
+        const from_str = try std.fmt.bufPrint(&from_buf, "{f}", .{from});
+        const to_str = try std.fmt.bufPrint(&to_buf, "{f}", .{to});
 
         // TwelveData's max outputsize is 5000 data points per request.
         // For daily candles this covers ~20 years of trading days (~252/year),

src/tui.zig

@@ -468,6 +468,19 @@ pub const App = struct {
     /// observation engine async dispatch, etc.) has a wall-clock-
     /// independent tick to react to.
     frame: u64 = 0,
+    /// True when the active tab has async work in flight that needs
+    /// poll-driven redraws (e.g. review's observation panel with
+    /// pending async checks). Set by the tab (loadData / activate);
+    /// cleared by the tab when the work completes or the tab
+    /// deactivates. While set, the event handler keeps a one-shot
+    /// vxfw Tick timer armed so the draw loop wakes up to poll —
+    /// without it, vxfw sleeps until the next user event and async
+    /// results would sit invisible until a keypress.
+    wants_poll_tick: bool = false,
+    /// True while a vxfw Tick command is scheduled and not yet
+    /// delivered. Prevents stacking duplicate timers (each event
+    /// would otherwise arm another).
+    poll_tick_armed: bool = false,
 
     has_explicit_symbol: bool = false, // true if -s was used
 
@@ -519,6 +532,27 @@ pub const App = struct {
 
     fn typeErasedEventHandler(ptr: *anyopaque, ctx: *vaxis.vxfw.EventContext, event: vaxis.vxfw.Event) anyerror!void {
         const self: *App = @ptrCast(@alignCast(ptr));
+        // Poll-driven redraw support: while a tab has async work
+        // in flight (`wants_poll_tick`), keep a one-shot vxfw
+        // Tick timer armed. The timer fires a `.tick` event,
+        // which requests a redraw; the draw pass runs the active
+        // tab's `tick` hook, which polls the async work and
+        // clears `wants_poll_tick` when done. Re-armed here on
+        // every event (including the .tick itself) until the tab
+        // clears the flag. 100ms cadence — fast enough to feel
+        // responsive, slow enough to be invisible in CPU terms.
+        defer if (self.wants_poll_tick and !self.poll_tick_armed) {
+            if (ctx.tick(100, self.widget())) {
+                self.poll_tick_armed = true;
+            } else |err| {
+                // OOM appending the Tick command. Leave
+                // poll_tick_armed false so the next event retries
+                // the arm; worst case the poll degrades to
+                // event-driven (pre-timer behavior), not a
+                // permanently stuck spinner.
+                std.log.scoped(.tui).warn("poll tick arm failed: {t}", .{err});
+            }
+        };
         switch (event) {
             .key_press => |key| {
                 // Tab-level pre-empt. The active tab gets first
@@ -553,6 +587,16 @@ pub const App = struct {
             .init => {
                 self.loadTabData();
             },
+            .tick => {
+                // Our scheduled poll timer fired. Mark it
+                // un-armed (the defer above re-arms if the tab
+                // still wants polling) and request a redraw so
+                // the draw pass runs the tab's tick hook.
+                self.poll_tick_armed = false;
+                if (self.wants_poll_tick) {
+                    ctx.redraw = true;
+                }
+            },
             else => {},
         }
     }

src/tui/review_tab.zig

@@ -142,6 +142,13 @@ pub const State = struct {
     /// Whether already-acknowledged findings are rendered. Toggled
     /// by `toggle_show_acked` (`v` key by default).
     show_acked: bool = false,
+    /// True while the observation panel has unresolved async
+    /// checks. Set when `loadData` builds a view whose panel
+    /// isn't complete; cleared by `tick` once every check
+    /// resolves (at which point findings_view is rebuilt to
+    /// include the late results). Keeps the per-frame tick poll
+    /// cost at zero for the common all-sync case.
+    observations_pending: bool = false,
 
     // ── Single-cursor (single-sheet) navigation ──────────────
 
@@ -253,6 +260,10 @@ pub const tab = struct {
 
     pub fn activate(state: *State, app: *App) !void {
         if (tab.isDisabled(app)) return;
+        // Re-arm the poll timer if we still have async checks in
+        // flight from a previous activation (user switched away
+        // and back while a check was running).
+        if (state.observations_pending) app.wants_poll_tick = true;
         if (state.loaded) return;
         // Load the journal before building findings; `loadData`
         // joins both into `findings_view`.
@@ -260,7 +271,16 @@ pub const tab = struct {
         loadData(state, app);
     }
 
-    pub const deactivate = framework.noopDeactivate(State);
+    /// Pause poll-driven redraws while this tab is inactive. The
+    /// tick hook only dispatches to the ACTIVE tab, so leaving
+    /// `wants_poll_tick` set after switching away would spin the
+    /// 100ms timer with nobody polling. `activate` re-arms when
+    /// the user returns; the async check keeps running in the
+    /// background either way.
+    pub fn deactivate(state: *State, app: *App) void {
+        _ = state;
+        app.wants_poll_tick = false;
+    }
 
     /// Manual refresh: drops the cached view and re-builds. Also
     /// invalidates the shared `account_map` and `classification_map`
@@ -280,7 +300,29 @@ pub const tab = struct {
         loadData(state, app);
     }
 
-    pub const tick = framework.noopTick(State);
+    /// Poll in-flight async observation checks. No-op (one bool
+    /// test) unless `loadData` left the panel incomplete. When
+    /// the last pending check resolves, rebuild the findings
+    /// view so late findings appear, and clear both the local
+    /// flag and the App-level poll-tick request so subsequent
+    /// frames cost nothing.
+    ///
+    /// Redraw cadence: while `app.wants_poll_tick` is set, the
+    /// App keeps a 100ms vxfw Tick timer armed (see
+    /// `typeErasedEventHandler`), so this hook gets called
+    /// even when the user isn't generating events. Without
+    /// that timer, async results would sit invisible until
+    /// the next keypress.
+    pub fn tick(state: *State, app: *App, frame: u64) void {
+        _ = frame;
+        if (!state.observations_pending) return;
+        const view = if (state.view) |*v| v else return;
+        const panel = if (view.observations) |*p| p else return;
+        if (!panel.isComplete()) return;
+        state.observations_pending = false;
+        app.wants_poll_tick = false;
+        rebuildFindingsView(state, app);
+    }
 
     pub fn handleAction(state: *State, app: *App, action: Action) void {
         switch (action) {
@@ -954,6 +996,20 @@ fn loadData(state: *State, app: *App) void {
 
     applySort(state);
     rebuildFindingsView(state, app);
+
+    // If any check is still running (async dispatch), arm the
+    // tick-poll so the findings view refreshes when results
+    // land. `isComplete` also transitions newly-finished checks,
+    // so a fast async check that finished during buildReview is
+    // caught right here without waiting for the next frame.
+    // `wants_poll_tick` tells the App to keep a vxfw Tick timer
+    // armed so the poll runs even with no user input.
+    if (state.view) |*v| {
+        if (v.observations) |*panel| {
+            state.observations_pending = !panel.isComplete();
+            if (state.observations_pending) app.wants_poll_tick = true;
+        }
+    }
 }
 
 /// Resolve the absolute path to `acknowledgments.srf` next to the
@@ -1462,11 +1518,15 @@ fn appendStatusGrid(
         // Promote the row's worst-severity color so multi-cell rows
         // with mixed states still draw the user's eye to the bad
         // ones. Order from best to worst: pass < skipped < warn <
-        // flag/err.
+        // flag/err. Pending checks rank with pass (no news isn't
+        // bad news yet).
         var row_style = th.mutedStyle();
-        var worst: u8 = 0; // 0=pass/skipped, 1=warn, 2=flag/err
+        var worst: u8 = 0; // 0=pass/skipped/pending, 1=warn, 2=flag/err
         for (panel.pending[i..end]) |pc| {
-            const result = pc.state.complete;
+            const result = switch (pc.state) {
+                .complete => |r| r,
+                .pending => continue,
+            };
             const rank: u8 = switch (result) {
                 .pass, .skipped => 0,
                 .warn => 1,
@@ -1482,7 +1542,11 @@ fn appendStatusGrid(
         try text.appendSlice(arena, "  ");
         for (panel.pending[i..end], 0..) |pc, col| {
             if (col > 0) try text.appendSlice(arena, "  ");
-            try appendStatusCell(arena, &text, pc.check.label, pc.state.complete);
+            const result: ?observations.CheckResult = switch (pc.state) {
+                .complete => |r| r,
+                .pending => null,
+            };
+            try appendStatusCell(arena, &text, pc.check.label, result);
         }
         try lines.append(arena, .{ .text = try text.toOwnedSlice(arena), .style = row_style });
 
@@ -1492,12 +1556,14 @@ fn appendStatusGrid(
 
 /// Append one cell's bytes to `text`: right-padded label + space
 /// + glyph. With ASCII glyphs (always 1 col), the cell ends at
-/// exactly `status_cell_cols` display columns.
+/// exactly `status_cell_cols` display columns. A `null` result
+/// means the check is still running (async dispatch in flight);
+/// rendered as an hourglass.
 fn appendStatusCell(
     arena: std.mem.Allocator,
     text: *std.ArrayList(u8),
     label: []const u8,
-    result: observations.CheckResult,
+    result: ?observations.CheckResult,
 ) !void {
     const lbl_cols = label.len; // ASCII labels: byte count == display cols
     if (lbl_cols < status_label_cols) {
@@ -1505,7 +1571,11 @@ fn appendStatusCell(
     }
     try text.appendSlice(arena, label);
     try text.append(arena, ' ');
-    try text.appendSlice(arena, checkStatusGlyph(result));
+    if (result) |r| {
+        try text.appendSlice(arena, checkStatusGlyph(r));
+    } else {
+        try text.appendSlice(arena, "⏳\u{FE0F}");
+    }
 }
 
 /// Append the findings section to `lines`. Layout:
@@ -3102,6 +3172,7 @@ test "appendStatusGrid: one row per status_cells_per_row checks" {
     };
     const panel: observations.CheckPanel = .{
         .allocator = testing.allocator,
+        .io = std.testing.io,
         .pending = &pending,
     };
 
@@ -3128,6 +3199,7 @@ test "appendStatusGrid: empty panel produces no lines" {
 
     const panel: observations.CheckPanel = .{
         .allocator = testing.allocator,
+        .io = std.testing.io,
         .pending = &.{},
     };
     var lines: std.ArrayList(StyledLine) = .empty;

src/views/observations_view.zig

@@ -102,6 +102,10 @@ pub fn build(
                 .flag => |o| o,
                 .pass, .skipped, .err => continue,
             },
+            // Still running — no findings to show yet. The caller
+            // rebuilds the view when the panel completes (TUI
+            // polls via tick; CLI awaits before building).
+            .pending => continue,
         };
 
         for (obs_slice) |obs| {
@@ -208,7 +212,7 @@ fn makePanel(allocator: std.mem.Allocator, obs_slice: []const Observation) !Chec
         .check = &check_singleton.c,
         .state = .{ .complete = .{ .warn = owned_obs } },
     };
-    return .{ .allocator = allocator, .pending = pending };
+    return .{ .allocator = allocator, .io = std.testing.io, .pending = pending };
 }
 
 fn makeJournalWithAck(
@@ -374,7 +378,7 @@ test "build: pass/skipped/err checks contribute no rows" {
         .check = &check_singleton.c,
         .state = .{ .complete = .pass },
     };
-    var panel = CheckPanel{ .allocator = testing.allocator, .pending = pending };
+    var panel = CheckPanel{ .allocator = testing.allocator, .io = std.testing.io, .pending = pending };
     defer panel.deinit();
 
     var journal = try makeEmptyJournal(testing.allocator);