v1.0.0-rc.7

Changed

• run's stderr is grouped into bold-cyan Overview / Warnings / Results / Summary
  sections in the same key value columns as plan: scan, transport with its fidelity
  coloured by trust, scope, timing, scan id + seed; warnings in their own block; a dim
  column header under Results on a terminal; the closing counts as result / states
  (each state in its colour) / probed / record. Warnings keep their wording.

Added

• docs/tutorial/demos/: an asciinema cast per tutorial section, linked from each use case; demos/record re-records them against the lab.

v1.0.0-rc.6

Added

• Release binaries for many more architectures, cross-compiled with cargo-zigbuild
  (no VMs — scanr is pure Rust with no C deps): static-musl aarch64, armv7, i686,
  riscv64gc and powerpc64le, and glibc aarch64, alongside the existing
  x86_64 gnu+musl. scripts/build-all.sh builds the whole set locally. Linux x86_64
  stays the tested platform; the rest are built and smoke-run under emulation,
  best-effort (docs/stability.md).

Changed

• plan groups its rows into coloured sections — Overview, Transport, Scope, Timing,
  Probing, Output, Projection, Host — for fast visual inspection. Section headers are
  bold cyan; fidelity is coloured by trust (green full, yellow open_only, red unknown);
  an armed TLS probe and the timeout-bound projection are highlighted; provenance stays
  dim. The two projections are labelled rate-bound and timeout-bound. Colour follows
  the usual rules (a terminal, no --no-color/NO_COLOR); piped output is unchanged
  plain text. plan rendering is not a promised surface.
• The tutorial lab is a single installable uv script, docs/tutorial/scanr-lab
  (PEP 723, typer), replacing the bash docs/tutorial/lab. up prints an aligned
  summary of the services and proxies it owns (no longer echoing lab.py's banner with
  its misleading "Ctrl-C to stop"); check shows a live up / down per port including
  the tunnel; down also sweeps an ssh -D orphaned by a stale pidfile. install /
  uninstall symlink it into uv tool dir --bin so scanr-lab runs from anywhere.
  lab.py stays plain Python — tests/tutorial.rs runs it directly, so the test path
  and CI are unchanged.

v1.0.0-rc.5

Changed

• Record filenames are scan-<name>-<UTC time>-<id>.jsonl.gz, e.g.
  scan-internal_web-2026_08_27T16_45_46Z-fafcc832.jsonl.gz
  (scan-adhoc-… for an ad-hoc run). Was scan-<epoch-ms>-<id>. The scan name (from
  config; adhoc when none) makes a directory of records self-describing, and the epoch
  is now a readable YYYY_MM_DDThh_mm_ssZ UTC stamp that still sorts chronologically.
  - is the sole field separator and appears in no field: a hyphen in the scan name is
  slugified to _ (internal-web → internal_web) and the stamp splits date from time
  with T, so basename.split('-') yields exactly [scan, name, stamp, id]. The
  scan- prefix stays, so scan-* globs and find_record are unaffected; the record's
  own scan_name/scan_id/started_at_epoch_ms remain the authority. Filenames are not
  a promised surface.
• RELEASING.md records the release cadence: batch small non-surface changes into
  [Unreleased] rather than cutting a release candidate for each; tag when a promised
  surface moves or a batch is worth a build.

v1.0.0-rc.4

Changed

• plan's dns row names the effective mode's resolver, not just the mode:
  auto -> local (names resolved on this host), transport (names handed to the proxy to resolve), disabled (hostname targets rejected). A pinned mode shows only itself;
  auto shows both the request and what it resolved to, so it never leaves the reader
  guessing which resolver a scan used.

v1.0.0-rc.3

Added

• scan_config.tls now records the hello's offer, not just the versions:
  offered_ciphers (suite names, the 1.3 suite first), offered_alpn, offered_groups
  and offered_sigalgs. The offer is fixed for every probe, so it is recorded once in
  the config event; a record now states both what was offered and what each server
  chose, without cross-referencing docs/security.md. With --tls-versions,
  version_hellos becomes one object per survey version (ssl2, ssl3, 1.0, 1.1,
  1.2) carrying that hello's sent_bytes and the era-appropriate ciphers it offers,
  in place of the bare sizes.

v1.0.0-rc.2

Added

• docs/tutorial.md, a learn-by-using guide with real output, and docs/tutorial/, its
  lab: ./lab up|tunnel|check|down brings up three loopback services (lab.py,
  standard library), squid/tinyproxy/3proxy/dante in rootless podman, and an ssh -D
  through a throwaway sshd. tests/tutorial.rs parses every scanr command in the
  document against the CLI and re-runs its use cases — direct ones against the lab,
  proxied ones against the fixtures — so the guide cannot rot.

Fixed

• output results shows the banner and the TLS summary as a trailing column of its
  table, exactly as run printed them live; replaying a record no longer hides the
  evidence it was kept for. --full, on run and on output results, shows banners
  untruncated (the default cuts at 48 characters; the printable-ASCII filter applies
  either way).
• The TLS probe reads the leaf certificate. tls.cert carries subject, issuer,
  alternative names, validity at probe time (valid, expired, not_yet_valid),
  whether issuer equals subject (self_signed) and the key type; result lines and
  output results say cn=host self-signed expired — for records written before this
  too, from the DER they already hold. Read, never verified; bounded and fuzzed
  (x509_leaf). The leaf also carries serial, sig_alg and version; a SHA-1 or
  MD5 signature is flagged sha1-signed / md5-signed on the line.
• The TLS flight is read to ServerHelloDone rather than stopping at the leaf, so the
  record gains chain (the certificates after the leaf, hashed and read, up to 8), the
  ECDHE kx_group and sig_scheme from ServerKeyExchange, compression, and the
  ServerHello's server_extensions with secure_renegotiation, extended_master_secret
  and session_ticket derived from it. negotiated says ssl3 for an SSLv3 answer.
• The TLS probe reads TLS 1.3 servers. The hello (now 218 bytes) offers 1.3 and 1.2;
  when a server takes 1.3 the probe finishes the key exchange with a published X25519
  key and decrypts the flight — certificate, chain, ALPN, signature scheme — instead of
  recording a protocol_version alert. offered says "1.3,1.2"; a HelloRetryRequest
  is recorded as hello_retry_request and not pursued. X25519, HKDF and AES-128-GCM are
  hand-rolled in src/crypto.rs, each held to an RFC or NIST vector; nothing is sent
  after the hello.
• --tls-versions (tls_versions = true, needs tls): after the hello, ask SSLv2,
  SSLv3, TLS 1.0, 1.1 and — when the server took 1.3 — 1.2 for themselves, each on its
  own connection with a hello of its era. The record's tls.versions says which are
  accepted, legacy_only when the newest is older than 1.2, and advice naming a client
  that can still reach it; the result line ends versions:ssl3..1.3 or
  legacy-only:tls1.0. An SSLv2-only server's certificate is taken from its
  SERVER-HELLO. Up to five more connections per silent open port; scan_config.tls
  lists the hellos' sizes and docs/security.md their bytes.
• SNI travels on the direct path when the target was given as a name: a locally resolved
  address now keeps its hostname, so a virtual-hosting server answers with its certificate
  instead of an alert. A real 443 answered internal_error to the nameless hello.
• output results --format json now carries banner, banner_hex, banner_bytes and
  tls when the record has them. The reader built for handing results to other tools
  dropped exactly the evidence those tools want; events always had it.

Added

• scanr plan projects the timeout-bound duration beside the rate-bound one: ~3d21h17m if every probe times out (5s x 2 attempts / 512 in flight). The rate cap never binds on a
  network of silent ports; this is the number that does.

Changed

• Direct scans are 2.5–4× faster. Workers hand results to the collector in batches
  of up to 64 (an open or a pressure event still flushes at once), and the hot path no
  longer allocates per probe: target names are formatted once, reasons are borrowed,
  attempt states live inline, the TLS observation is boxed. Measured on a 64-core box,
  loopback /24, every port refused: 256k probes 1.54 s → 0.40 s at the default
  concurrency (166k → 640k probes/s); 2.56M probes 14.1 s → 4.3 s. Throughput no longer
  falls with concurrency (512 was 40% slower than 64; now within 5%). Record contents,
  order and every reader output are unchanged — the compat corpus passes as pinned.

v1.0.0-rc.1

Added

• docs/stability.md: what 1.x promises (record, CLI, config), what it does not, and
  the deprecation rule.
• tests/compat/: eleven scenario records — direct (spans and rows), SOCKS5 full and
  open-only, HTTP CONNECT, a mixed chain, a pool, banner + TLS, interrupted and resumed,
  and a schema-1 record written by the 0.2.2 binary — with the exact output of every
  reader pinned, plus the config init template of every release, which must still
  validate and plan. tests/compat.rs holds the build to all of it.
• docs/evidence.md maps every claim in the docs to the test, corpus scenario or
  measured decision behind it; tests/evidence.rs fails if a cited test disappears.
• The nmap differential now covers HTTP CONNECT, and CI checks the TLS probe against
  openssl s_server.

Changed

• output summarize takes --format table|json, the same flag as output results.

Deprecated

• output summarize --json: still works, warns on stderr, removed in 2.0.

v0.5.0

Added

• The TLS ClientHello probe, --tls / tls = true — off by default, the one active
  probe scanr has. On open ports that volunteered no banner it sends a fixed 163-byte TLS
  1.2 ClientHello on the same connection and records what comes back: negotiated version,
  cipher, ALPN, the leaf certificate (SHA-256, and the DER when ≤ 8 KiB) and chain length,
  or the alert — a 1.3-only server answers protocol_version. Verified against openssl s_server. scan_config.tls.sent_bytes states what was sent (0 when off);
  docs/security.md lists the exact bytes and a test holds it to them. tls_timeout (1s)
  is the profile ceiling. Fuzz target tls_reply. No new dependencies.

v0.4.0

Added

• HTTP CONNECT proxies: type = "http" with the same keys as socks5; Basic
  authentication; usable as a chain hop (mixed with socks5 in either order) and a pool
  member. Measured against squid 7.6, tinyproxy 1.11.2 and 3proxy 0.9.7: none has a
  status meaning "refused" (503/500/502, the same for a timeout), so an http
  transport is open_only by construction — non-open results are error carrying the
  status line, fidelity_source is inherent, and declaring fidelity = "full" is
  refused. Fuzz target http_connect_reply covers the response parser.

Changed

• A chain's fidelity is its exit hop's, not its weakest hop's. Only the last CONNECT
  names the destination and its reply travels back untouched; measured full through
  squid → 3proxy SOCKS5. fidelity_source for a chain is exit_hop (was weakest_hop);
  chain hops[] and pool members[] entries now carry type.
• transport test prints status NNN for HTTP replies beside reply 0xNN for SOCKS5.

Fixed

• A chain or pool resolved hostnames locally and took the direct default profile.
  supports_remote_dns was true only for a single socks5, so dns = "auto" through a
  chain leaked DNS from the host and the scan ran with no rate limit. Every proxied kind
  now resolves through the transport; a pool does when every member does.
• Chains and pools with unknown or open_only fidelity now get the fidelity_unknown
  / fidelity_open_only plan warnings; only a single socks5 did.

v0.3.0

Changed

• schema_version is 2: probe_span ranges are counter indices, not matrix positions, so
  collapse no longer decays with scan duration (10,023 → 595 ranges, 53,765 → 4,893 B on a
  rate-limited 20,001-probe scan). Expansion needs the recorded seed. Writes 2, reads 1 and 2.
• docs/output-schema.md states which enumerations are closed (state, source) and which
  are open (transport.type, fidelity fields, warning and error codes); a test pins the closed sets.
• Docs compacted; docs/design/decisions.md is a terse register; ROADMAP.md added with
  the 1.0 policy and plan.

Security

• The record is created mode 0600, including the .partial file.
• password_file mode is checked on the descriptor it is read from.
• All 31 GitHub Actions references pinned to commit SHAs; contents: write scoped to the
  publish job; CI declares contents: read.
• unsafe_code denied crate-wide; each of the five blocks carries #[allow] and a safety
  comment; docs/security.md lists them and a test checks both directions.
• Threat model covers chains: every hop sees the credentials of every hop after it.
• build.rs no longer stamps an enclosing repository's commit when the tree has no .git.
• deny.toml multiple-versions = "deny", with the one known duplicate listed.

Fixed

• output verify exits 2 for a bad record and 1 for an unreadable file.
• output results --hosts <name> says on stderr when a name cannot match an address record.
• config init documents all keys (banner, compress, spans, services_file,
  use_etc_services, banner_bytes, banner_timeout, hops, members), seven profiles,
  and four transport types; drift guards are exhaustive destructurings.
• transport list no longer prints direct twice when a config redefines it.
• docs/cli.md guards every subcommand's flags, not only run's.
• output results takes --format json, not --json (docs corrected).
• RawDefaults::banner documented as on by default.
• Usage errors exit 1, not 2; --help and --version exit 0.
• The record's exit_code reflects writer failures.
• output verify accepts every schema version the build can read.
• A version-2 record with a missing or unparseable seed is reported, not expanded wrongly.