fix: clear inspect child environment by default

[nightcityblade]

What & why

Make the inspect example clear the child environment by default so it matches hermetic termlens tests. PATH is deliberately retained so bare program names still resolve; repeatable --env KEY=VALUE adds selected variables, while --inherit-env restores the previous behavior. End-to-end coverage verifies both clearing and selective injection.

Closes #263.

Checklist

• Linked an issue (or explained above why none exists)
• Tests added/updated for the change
• cargo fmt --all and cargo clippy --workspace --all-targets --all-features are clean
• All commits are signed off (git commit -s) — see CONTRIBUTING.md §5
• No AI attribution trailers (no AI co-authors, "Generated with" footers, or bot identities) — see CONTRIBUTING.md §6
• CHANGELOG.md updated under [Unreleased] (user-facing changes only)
• Snapshot changes (if any) were reviewed with cargo insta review, not blind-accepted (N/A: no snapshots changed)

[vyncint]

Reviewed and verified. The behaviour is right and the PATH decision is the one I would have made.

What I checked beyond CI, by running the built example:

• Default clears the environment: a FOO=leaked in the caller's shell reaches the child as unset, while PATH survives.
• --env is genuinely repeatable (--env A=1 --env B=2 sets both), splits on the first = so --env 'K=a=b' gives K the value a=b, and rejects both a missing = and an empty key in one line with exit 1, in the same shape --size uses.
• --inherit-env restores the old behaviour exactly.
• A bare program name still resolves, which is the whole point of keeping PATH — and your new test exercises that path by spawning sh, which is neat.
• --help lists both new flags, and the usage text stays one source.

Keeping PATH is a documented exception rather than a hole, and the usage text says so, which is what the issue asked for.

Two more changes, alongside the shared one below

1. The new test's assertion has no failure message. Every other assertion in crates/termlens/tests/inspect.rs prints the output it was reading — "terminal size missing from:\n{stdout}" and so on — so a failure shows what the screen actually held. As written, a regression here prints assertion failed: …contains("unset:yes") and nothing else. Please give it the same treatment.

2. --inherit-env has no test. It is the flag that restores the previous behaviour, so it is the one most likely to be quietly broken by a later change to the argument loop. One more case in the same test, asserting that a variable set on the Command does reach the child under --inherit-env, would close it. That also makes the pair of cases read as one statement about the policy.

---

One thing to fix, and it is the same in all five of your PRs

The CHANGELOG entry landed inside the published ## [0.9.0] - 2026-09-05 section rather than under ## [Unreleased]. Easy to miss: [Unreleased] was empty when you branched, so the nearest heading with content under it belongs to 0.9.0.

It matters for two reasons. 0.9.0 is already on crates.io and its GitHub Release notes were generated from that section, so the entry now describes a release that does not contain the change. And docs/RELEASING.md cuts the next release by moving [Unreleased] down into a new version heading — an entry parked in [0.9.0] would never appear in any release notes at all.

Move it up under ## [Unreleased], adding the ### Added / ### Changed / ### Fixed sub-heading there if it does not exist yet. main has moved on by a commit since you branched, so a rebase is wanted anyway:

git fetch origin && git rebase origin/main

Thanks for these — five well-scoped fixes with tests and docs, and the PR bodies say what you decided and why. That is exactly the shape this repository asks for.

[nightcityblade]

Addressed all requested changes in 5d3aec5: rebased and moved the changelog entry under Unreleased → Changed, added assertion diagnostics, and added coverage for --inherit-env.

Validation: cargo fmt --check and cargo test --workspace --all-features pass.

[vyncint]

Thanks — all three points addressed. Both assertions now print what they read, and the --inherit-env case is covered, which is the flag most likely to be broken quietly by a later change to the argument loop.

Re-verified on this head by running the built example: the default clears the environment while keeping PATH, --env is repeatable and splits on the first =, a malformed value is one line and exit 1, --inherit-env restores the old behaviour, and a bare program name still resolves.

I rebased onto main and resolved the changelog collision with the three that landed ahead of this. Your commit, author and sign-off are untouched. Merging — and thank you for five well-scoped PRs in a day, each with the reasoning written down.