[vjovanov]

Closes #244

After a separate ephor refresh demo successfully replaced OLD cache row with NEW cache row on disk, an open ephor tui used to keep showing OLD without saying the screen was behind. It now keeps that displayed snapshot and adds a persistent warning on the periodic tick:

Displayed feed outdated: demo — newer cache available; press r to refresh

The warning names every affected project and wraps to the terminal width. Rows, selection and seen/done state stay intact. Other messages and unrelated project refreshes cannot clear it: each project's warning disappears only when an existing feed load or refresh absorption brings that displayed snapshot up to the newer cache. The new functional point specifies this behavior and keeps it separate from provider-failure stale labeling (§FS-001-forge-interface.7.1).

The TUI observes only its session projects' known cache files. Unchanged metadata avoids parsing and added redraws; changed files are checked for valid newer fetch evidence. Initial loading, reloads and per-project absorption establish their metadata baseline before reading the displayed snapshot, so a replacement during loading remains detectable. The updated session explanation documents this local observation and when warnings retire (§AR-009-surfaces.2).

Observation starts no provider fetch, changes no provider health and writes no cache. Missing, unreadable, malformed, incompatible or wrong-project cache data preserves last-good rows and any established warning. The existing r action remains the remedy; CLI/JSON output, cache schema, provider interfaces and refresh scope retain their contracts.

The specification and three failing regressions were committed first in 86d318c6db; implementation 6f4b671dba enabled all three without changing their assertions. The real-terminal regression uses pyte==0.8.2, added as a test dependency and installed by CI before Cargo tests.

Local verification on 6f4b671dba passed:

• All three regressions: real-terminal external refresh and r recovery, rendered warning lifecycle with reader-state preservation, and deterministic cache replacement during the initial read. None is ignored.
• The unchanged triage reproducer: OLD remained on the current screen alongside the exact demo warning, NEW was on disk, and only the initial and external forge fetches occurred. Exit 0.
• pre-commit run --all-files: Rust formatting, warnings-as-errors build, all-target Rust tests, Python integration, CI-pinned grund 0.18.0 check/format, fissile and content guards.
• Standalone Python integration discovery: 113 tests passed; boundary and parity checks passed; explicit pinned grund fmt --check passed.
• The complete Rust suite passed again with a symlinked TMPDIR, including all three regressions.

The local gate finished READY: yes on its first run; independent review approved with no findings. Remote CI is checked during shipping. The generated workflow summary below covers recorded invocations before shipping.

AI workflow: `rhei`, 11 agent invocations across 1 model; 6 tasks completed, 10 in progress.

1. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 1) — cdx, openai/gpt-6.1-sol — 2m22s — 555.6k in / 4.6k out
2. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket.triage assessing (visit 1) — cdx, openai/gpt-6.1-sol — 1m32s — 303.7k in / 2.5k out
3. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket.triage reproducing (visit 1) — cdx, openai/gpt-6.1-sol — 7m42s — 1.3M in / 14.4k out
4. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 2) — cdx, openai/gpt-6.1-sol — 2m34s — 707.3k in / 6.8k out
5. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket.specify specify — cdx, openai/gpt-6.1-sol — 15m03s — 4.3M in / 28.8k out
6. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 3) — cdx, openai/gpt-6.1-sol — 1m56s — 455.6k in / 4.0k out
7. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket.implement implement — cdx, openai/gpt-6.1-sol — 10m19s — 2.8M in / 16.8k out
8. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 4) — cdx, openai/gpt-6.1-sol — 2m24s — 450.6k in / 4.2k out
9. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket.review-1 review — cdx, openai/gpt-6.1-sol — 5m17s — 1.3M in / 9.6k out
10. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 5) — cdx, openai/gpt-6.1-sol — 2m45s — 824.1k in / 4.7k out
11. github-issues-agent-grounds-ephor-244-implement-8b944eef.ticket supervising (visit 6) — cdx, openai/gpt-6.1-sol — 2m07s — 514.8k in / 4.5k out

Accounting	Value
cost	$4.17
total tokens	13.6M
input tokens (incl. cache)	13.5M
input cache read	12.6M
input cache write	-
output tokens (incl. cache)	101.0k
output cache read	-
output cache write	-
coverage	Complete